> ## Documentation Index
> Fetch the complete documentation index at: https://docs2.openclaw.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Configuration — agents

Agent-scoped configuration keys under `agents.*`, `multiAgent.*`, `session.*`,
`messages.*`, and `talk.*`. For channels, tools, gateway runtime, and other
top-level keys, see [Configuration reference](/gateway/configuration-reference).

## Agent defaults

### `agents.defaults.workspace`

Default: `OPENCLAW_WORKSPACE_DIR` when set, otherwise `~/.openclaw/workspace` (or `~/.openclaw/workspace-<profile>` when `OPENCLAW_PROFILE` is set to a non-default profile).

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { workspace: "~/.openclaw/workspace" } },
}
```

An explicit `agents.defaults.workspace` value takes precedence over
`OPENCLAW_WORKSPACE_DIR`. Use the environment variable to point default agents
at a mounted workspace when you do not want to write that path into config.

### `agents.defaults.repoRoot`

Optional repository root shown in the system prompt's Runtime line. If unset, OpenClaw auto-detects by walking upward from the workspace.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { repoRoot: "~/Projects/openclaw" } },
}
```

### `agents.defaults.skills`

Optional default skill allowlist for agents that do not set
`agents.list[].skills`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: { skills: ["github", "weather"] },
    list: [
      { id: "writer" }, // inherits github, weather
      { id: "docs", skills: ["docs-search"] }, // replaces defaults
      { id: "locked-down", skills: [] }, // no skills
    ],
  },
}
```

* Omit `agents.defaults.skills` for unrestricted skills by default.
* Omit `agents.list[].skills` to inherit the defaults.
* Set `agents.list[].skills: []` for no skills.
* A non-empty `agents.list[].skills` list is the final set for that agent; it
  does not merge with defaults.

### `agents.defaults.skipBootstrap`

Disables automatic creation of workspace bootstrap files (`AGENTS.md`, `SOUL.md`, `TOOLS.md`, `IDENTITY.md`, `USER.md`, `HEARTBEAT.md`, `BOOTSTRAP.md`).

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { skipBootstrap: true } },
}
```

### `agents.defaults.skipOptionalBootstrapFiles`

Skips creation of selected optional workspace files while still writing required bootstrap files (`AGENTS.md`, `TOOLS.md`, `BOOTSTRAP.md`). Valid values: `SOUL.md`, `USER.md`, `HEARTBEAT.md`, and `IDENTITY.md`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      skipOptionalBootstrapFiles: ["SOUL.md", "USER.md"],
    },
  },
}
```

### `agents.defaults.contextInjection`

Controls when workspace bootstrap files are injected into the system prompt. Default: `"always"`.

* `"continuation-skip"`: safe continuation turns (after a completed assistant response) skip workspace bootstrap re-injection, reducing prompt size. Heartbeat runs and post-compaction retries still rebuild context.
* `"never"`: disable workspace bootstrap and context-file injection on every turn. Use this only for agents that fully own their prompt lifecycle (custom context engines, native runtimes that build their own context, or specialized bootstrap-free workflows). Heartbeat and compaction-recovery turns also skip injection.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { contextInjection: "continuation-skip" } },
}
```

Per-agent override: `agents.list[].contextInjection`. Omitted values inherit
`agents.defaults.contextInjection`.

### `agents.defaults.bootstrapMaxChars`

Max characters per workspace bootstrap file before truncation. Default: `20000`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { bootstrapMaxChars: 20000 } },
}
```

Per-agent override: `agents.list[].bootstrapMaxChars`. Omitted values inherit
`agents.defaults.bootstrapMaxChars`.

### `agents.defaults.bootstrapTotalMaxChars`

Max total characters injected across all workspace bootstrap files. Default: `60000`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { bootstrapTotalMaxChars: 60000 } },
}
```

Per-agent override: `agents.list[].bootstrapTotalMaxChars`. Omitted values
inherit `agents.defaults.bootstrapTotalMaxChars`.

### Per-agent bootstrap profile overrides

Use per-agent bootstrap profile overrides when one agent needs different prompt
injection behavior from the shared defaults. Omitted fields inherit from
`agents.defaults`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      contextInjection: "continuation-skip",
      bootstrapMaxChars: 20000,
      bootstrapTotalMaxChars: 60000,
    },
    list: [
      {
        id: "strict-worker",
        contextInjection: "always",
        bootstrapMaxChars: 50000,
        bootstrapTotalMaxChars: 300000,
      },
    ],
  },
}
```

### `agents.defaults.bootstrapPromptTruncationWarning`

Controls the agent-visible system-prompt notice when bootstrap context is truncated.
Default: `"always"`.

* `"off"`: never inject truncation notice text into the system prompt.
* `"once"`: inject a concise notice once per unique truncation signature.
* `"always"`: inject a concise notice on every run when truncation exists (recommended).

Detailed raw/injected counts and config tuning fields stay in diagnostics such
as context/status reports and logs; routine WebChat user/runtime context only
gets the concise recovery notice.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { bootstrapPromptTruncationWarning: "always" } }, // off | once | always
}
```

### Context budget ownership map

OpenClaw has multiple high-volume prompt/context budgets, and they are
intentionally split by subsystem instead of all flowing through one generic
knob.

| Budget                                                         | Covers                                                                                                                                                          |
| -------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `agents.defaults.bootstrapMaxChars` / `bootstrapTotalMaxChars` | Normal workspace bootstrap injection                                                                                                                            |
| `agents.defaults.startupContext.*`                             | One-shot reset/startup model-run prelude, including recent daily `memory/*.md` files. Bare chat `/new` and `/reset` are acknowledged without invoking the model |
| `skills.limits.*`                                              | The compact skills list injected into the system prompt                                                                                                         |
| `agents.defaults.contextLimits.*`                              | Bounded runtime excerpts and injected runtime-owned blocks                                                                                                      |
| `memory.qmd.limits.*`                                          | Indexed memory-search snippet and injection sizing                                                                                                              |

Matching per-agent overrides:

* `agents.list[].skillsLimits.maxSkillsPromptChars`
* `agents.list[].contextInjection`
* `agents.list[].bootstrapMaxChars`
* `agents.list[].bootstrapTotalMaxChars`
* `agents.list[].contextLimits.*`

#### `agents.defaults.startupContext`

Controls the first-turn startup prelude injected on reset/startup model runs.
Bare chat `/new` and `/reset` commands acknowledge the reset without invoking
the model, so they do not load this prelude.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      startupContext: {
        enabled: true,
        applyOn: ["new", "reset"],
        dailyMemoryDays: 2,
        maxFileBytes: 16384,
        maxFileChars: 1200,
        maxTotalChars: 2800,
      },
    },
  },
}
```

#### `agents.defaults.contextLimits`

Shared defaults for bounded runtime context surfaces.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      contextLimits: {
        memoryGetMaxChars: 12000,
        memoryGetDefaultLines: 120,
        postCompactionMaxChars: 1800,
      },
    },
  },
}
```

* `memoryGetMaxChars`: default `memory_get` excerpt cap before truncation
  metadata and continuation notice are added.
* `memoryGetDefaultLines`: default `memory_get` line window when `lines` is
  omitted.
* `toolResultMaxChars`: advanced live tool-result ceiling used for persisted
  results and overflow recovery. Leave unset for the model-context auto cap:
  `16000` chars below 100K tokens, `32000` chars at 100K+ tokens, and `64000`
  chars at 200K+ tokens. Explicit values up to `1000000` are accepted for
  long-context models, but the effective cap is still limited to about 30% of
  the model context window. `openclaw doctor --deep` prints the effective cap,
  and doctor warns only when an explicit override is stale or has no effect.
* `postCompactionMaxChars`: AGENTS.md excerpt cap used during post-compaction
  refresh injection.

#### `agents.list[].contextLimits`

Per-agent override for the shared `contextLimits` knobs. Omitted fields inherit
from `agents.defaults.contextLimits`.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      contextLimits: { memoryGetMaxChars: 12000 },
    },
    list: [
      {
        id: "tiny-local",
        contextLimits: {
          memoryGetMaxChars: 6000,
          toolResultMaxChars: 8000, // advanced ceiling for this agent
        },
      },
    ],
  },
}
```

#### `skills.limits.maxSkillsPromptChars`

Global cap for the compact skills list injected into the system prompt. This
does not affect reading `SKILL.md` files on demand.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  skills: { limits: { maxSkillsPromptChars: 18000 } },
}
```

#### `agents.list[].skillsLimits.maxSkillsPromptChars`

Per-agent override for the skills prompt budget.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    list: [{ id: "tiny-local", skillsLimits: { maxSkillsPromptChars: 6000 } }],
  },
}
```

### `agents.defaults.imageMaxDimensionPx`

Max pixel size for the longest image side in transcript/tool image blocks before provider calls.
Default: `1200`.

Lower values usually reduce vision-token usage and request payload size for screenshot-heavy runs.
Higher values preserve more visual detail.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { imageMaxDimensionPx: 1200 } },
}
```

### `agents.defaults.imageQuality`

Image-tool compression/detail preference for images loaded from file paths, URLs, and media references.
Default: `auto`.

OpenClaw adapts the resize ladder to the selected image model. For example, Claude Opus 4.8, OpenAI GPT-5.5, Qwen VL, and hosted Llama 4 vision models can use larger images than older/default high-detail vision paths, while multi-image turns are compressed more aggressively in `auto` mode to control token and latency cost.

Values:

* `auto`: adapt to model limits and image count.
* `efficient`: prefer smaller images for lower token and byte usage.
* `balanced`: use the standard middle-ground ladder.
* `high`: preserve more detail for screenshots, diagrams, and document images.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { imageQuality: "auto" } },
}
```

### `agents.defaults.userTimezone`

Timezone for system prompt context (not message timestamps). Falls back to host timezone.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { userTimezone: "America/Chicago" } },
}
```

### `agents.defaults.timeFormat`

Time format in system prompt. Default: `auto` (OS preference).

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: { defaults: { timeFormat: "auto" } }, // auto | 12 | 24
}
```

### `agents.defaults.model`

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      models: {
        "anthropic/claude-opus-4-6": { alias: "opus" },
        "minimax/MiniMax-M2.7": { alias: "minimax" },
      },
      model: {
        primary: "anthropic/claude-opus-4-6",
        fallbacks: ["minimax/MiniMax-M2.7"],
      },
      utilityModel: "openai/gpt-5.4-mini",
      imageModel: {
        primary: "openrouter/qwen/qwen-2.5-vl-72b-instruct:free",
        fallbacks: ["openrouter/google/gemini-2.0-flash-vision:free"],
      },
      imageGenerationModel: {
        primary: "openai/gpt-image-2",
        fallbacks: ["google/gemini-3.1-flash-image-preview"],
      },
      videoGenerationModel: {
        primary: "qwen/wan2.6-t2v",
        fallbacks: ["qwen/wan2.6-i2v"],
      },
      pdfModel: {
        primary: "anthropic/claude-opus-4-6",
        fallbacks: ["openai/gpt-5.4-mini"],
      },
      params: { cacheRetention: "long" }, // global default provider params
      pdfMaxBytesMb: 10,
      pdfMaxPages: 20,
      thinkingDefault: "low",
      verboseDefault: "off",
      toolProgressDetail: "explain",
      reasoningDefault: "off",
      elevatedDefault: "on",
      timeoutSeconds: 600,
      mediaMaxMb: 5,
      contextTokens: 200000,
      maxConcurrent: 4,
    },
  },
}
```

* `model`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * String form sets only the primary model.
  * Object form sets primary plus ordered failover models.
* `utilityModel`: optional `provider/model` ref or alias for short internal tasks. It currently powers generated Control UI session titles, Telegram DM topic titles, and Discord auto-thread titles. These tasks fall back to the agent's primary model when unset; `agents.list[].utilityModel` overrides the default, and an operation-specific model override wins over both. Utility tasks make separate model calls and send task-specific content to the selected model provider. Dashboard title generation sends at most the first 1,000 characters of the first non-command message. Choose a provider that matches your cost and data-handling requirements.
* `imageModel`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * Used by the `image` tool path as its vision-model config.
  * Also used as fallback routing when the selected/default model cannot accept image input.
  * Prefer explicit `provider/model` refs. Bare IDs are accepted for compatibility; if a bare ID uniquely matches a configured image-capable entry in `models.providers.*.models`, OpenClaw qualifies it to that provider. Ambiguous configured matches require an explicit provider prefix.
* `imageGenerationModel`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * Used by the shared image-generation capability and any future tool/plugin surface that generates images.
  * Typical values: `google/gemini-3.1-flash-image-preview` for native Gemini image generation, `fal/fal-ai/flux/dev` for fal, `openai/gpt-image-2` for OpenAI Images, or `openai/gpt-image-1.5` for transparent-background OpenAI PNG/WebP output.
  * If you select a provider/model directly, configure matching provider auth too (for example `GEMINI_API_KEY` or `GOOGLE_API_KEY` for `google/*`, `OPENAI_API_KEY` or OpenAI Codex OAuth for `openai/gpt-image-2` / `openai/gpt-image-1.5`, `FAL_KEY` for `fal/*`).
  * If omitted, `image_generate` can still infer an auth-backed provider default. It tries the current default provider first, then the remaining registered image-generation providers in provider-id order.
* `musicGenerationModel`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * Used by the shared music-generation capability and the built-in `music_generate` tool.
  * Typical values: `google/lyria-3-clip-preview`, `google/lyria-3-pro-preview`, or `minimax/music-2.6`.
  * If omitted, `music_generate` can still infer an auth-backed provider default. It tries the current default provider first, then the remaining registered music-generation providers in provider-id order.
  * If you select a provider/model directly, configure the matching provider auth/API key too.
* `videoGenerationModel`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * Used by the shared video-generation capability and the built-in `video_generate` tool.
  * Typical values: `qwen/wan2.6-t2v`, `qwen/wan2.6-i2v`, `qwen/wan2.6-r2v`, `qwen/wan2.6-r2v-flash`, or `qwen/wan2.7-r2v`.
  * If omitted, `video_generate` can still infer an auth-backed provider default. It tries the current default provider first, then the remaining registered video-generation providers in provider-id order.
  * If you select a provider/model directly, configure the matching provider auth/API key too.
  * The official Qwen video-generation plugin supports up to 1 output video, 1 input image, 4 input videos, 10 seconds duration, and provider-level `size`, `aspectRatio`, `resolution`, `audio`, and `watermark` options.
* `pdfModel`: accepts either a string (`"provider/model"`) or an object (`{ primary, fallbacks }`).
  * Used by the `pdf` tool for model routing.
  * If omitted, the PDF tool falls back to `imageModel`, then to the resolved session/default model.
* `pdfMaxBytesMb`: default PDF size limit for the `pdf` tool when `maxBytesMb` is not passed at call time.
* `pdfMaxPages`: default maximum pages considered by extraction fallback mode in the `pdf` tool.
* `verboseDefault`: default verbose level for agents. Values: `"off"`, `"on"`, `"full"`. Default: `"off"`.
* `toolProgressDetail`: detail mode for `/verbose` tool summaries and progress-draft tool lines. Values: `"explain"` (default, compact human labels) or `"raw"` (append raw command/detail when available). Per-agent `agents.list[].toolProgressDetail` overrides this default.
* `reasoningDefault`: default reasoning visibility for agents. Values: `"off"`, `"on"`, `"stream"`. Per-agent `agents.list[].reasoningDefault` overrides this default. Configured reasoning defaults are only applied for owners, authorized senders, or operator-admin gateway contexts when no per-message or session reasoning override is set.
* `elevatedDefault`: default elevated-output level for agents. Values: `"off"`, `"on"`, `"ask"`, `"full"`. Default: `"on"`.
* `model.primary`: format `provider/model` (e.g. `openai/gpt-5.5` for OpenAI API-key or Codex OAuth access). If you omit the provider, OpenClaw tries an alias first, then a unique configured-provider match for that exact model id, and only then falls back to the configured default provider (deprecated compatibility behavior, so prefer explicit `provider/model`). If that provider no longer exposes the configured default model, OpenClaw falls back to the first configured provider/model instead of surfacing a stale removed-provider default.
* `models`: the configured model catalog and allowlist for `/model`. Each entry can include `alias` (shortcut) and `params` (provider-specific, for example `temperature`, `maxTokens`, `cacheRetention`, `context1m`, `responsesServerCompaction`, `responsesCompactThreshold`, OpenRouter `provider` routing, `chat_template_kwargs`, `extra_body`/`extraBody`).
  * Use `provider/*` entries such as `"openai/*": {}` or `"vllm/*": {}` to show all discovered models for selected providers without manually listing every model id.
  * Add `agentRuntime` to a `provider/*` entry when every dynamically discovered model for that provider should use the same runtime. Exact `provider/model` runtime policy still wins over the wildcard.
  * Safe edits: use `openclaw config set agents.defaults.models '<json>' --strict-json --merge` to add entries. `config set` refuses replacements that would remove existing allowlist entries unless you pass `--replace`.
  * Provider-scoped configure/onboarding flows merge selected provider models into this map and preserve unrelated providers already configured.
  * For direct OpenAI Responses models, server-side compaction is enabled automatically. Use `params.responsesServerCompaction: false` to stop injecting `context_management`, or `params.responsesCompactThreshold` to override the threshold. See [OpenAI server-side compaction](/providers/openai#advanced-configuration).
* `params`: global default provider parameters applied to all models. Set at `agents.defaults.params` (e.g. `{ cacheRetention: "long" }`).
* `params` merge precedence (config): `agents.defaults.params` (global base) is overridden by `agents.defaults.models["provider/model"].params` (per-model), then `agents.list[].params` (matching agent id) overrides by key. See [Prompt Caching](/reference/prompt-caching) for details.
* `models.providers.openrouter.params.provider`: OpenRouter-wide default provider-routing policy. OpenClaw forwards this to OpenRouter's request `provider` object; per-model `agents.defaults.models["openrouter/<model>"].params.provider` and agent params override by key. See [OpenRouter provider routing](/providers/openrouter#advanced-configuration).
* `params.extra_body`/`params.extraBody`: advanced pass-through JSON merged into `api: "openai-completions"` request bodies for OpenAI-compatible proxies. If it collides with generated request keys, the extra body wins; non-native completions routes still strip OpenAI-only `store` afterward.
* `params.chat_template_kwargs`: vLLM/OpenAI-compatible chat-template arguments merged into top-level `api: "openai-completions"` request bodies. For `vllm/nemotron-3-*` with thinking off, the bundled vLLM plugin automatically sends `enable_thinking: false` and `force_nonempty_content: true`; explicit `chat_template_kwargs` override generated defaults, and `extra_body.chat_template_kwargs` still has final precedence. Configured vLLM Qwen and Nemotron thinking models expose binary `/think` choices (`off`, `on`) instead of the multi-level effort ladder.
* `compat.thinkingFormat`: OpenAI-compatible thinking payload style. Use `"together"` for Together-style `reasoning.enabled`, `"qwen"` for Qwen-style top-level `enable_thinking`, or `"qwen-chat-template"` for `chat_template_kwargs.enable_thinking` on Qwen-family backends that support request-level chat-template kwargs, such as vLLM. OpenClaw maps disabled thinking to `false` and enabled thinking to `true`, and configured vLLM Qwen models expose binary `/think` choices for these formats.
* `compat.supportedReasoningEfforts`: per-model OpenAI-compatible reasoning effort list. Include `"xhigh"` for custom endpoints that truly accept it; OpenClaw then exposes `/think xhigh` in command menus, Gateway session rows, session patch validation, agent CLI validation, and `llm-task` validation for that configured provider/model. Use `compat.reasoningEffortMap` when the backend wants a provider-specific value for a canonical level.
* `params.preserveThinking`: Z.AI-only opt-in for preserved thinking. When enabled and thinking is on, OpenClaw sends `thinking.clear_thinking: false` and replays prior `reasoning_content`; see [Z.AI thinking and preserved thinking](/providers/zai#advanced-configuration).
* `localService`: optional provider-level process manager for local/self-hosted model servers. When the selected model belongs to that provider, OpenClaw probes `healthUrl` (or `baseUrl + "/models"`), starts `command` with `args` if the endpoint is down, waits up to `readyTimeoutMs`, then sends the model request. `command` must be an absolute path. `idleStopMs: 0` keeps the process alive until OpenClaw exits; a positive value stops the OpenClaw-spawned process after that many idle milliseconds. See [Local model services](/gateway/local-model-services).
* Runtime policy belongs on providers or models, not on `agents.defaults`. Use `models.providers.<provider>.agentRuntime` for provider-wide rules or `agents.defaults.models["provider/model"].agentRuntime` / `agents.list[].models["provider/model"].agentRuntime` for model-specific rules. OpenAI agent models on the official OpenAI provider select Codex by default.
* Config writers that mutate these fields (for example `/models set`, `/models set-image`, and fallback add/remove commands) save canonical object form and preserve existing fallback lists when possible.
* `maxConcurrent`: max parallel agent runs across sessions (each session still serialized). Default: `4`.

### Runtime policy

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  models: {
    providers: {
      openai: {
        agentRuntime: { id: "codex" },
      },
    },
  },
  agents: {
    defaults: {
      model: "openai/gpt-5.5",
      models: {
        "anthropic/claude-opus-4-8": {
          agentRuntime: { id: "claude-cli" },
        },
        "vllm/*": {
          agentRuntime: { id: "openclaw" },
        },
      },
    },
  },
}
```

* `id`: `"auto"`, `"openclaw"`, a registered plugin harness id, or a supported CLI backend alias. The bundled Codex plugin registers `codex`; the bundled Anthropic plugin provides the `claude-cli` CLI backend.
* `id: "auto"` lets registered plugin harnesses claim supported turns and uses OpenClaw when no harness matches. An explicit plugin runtime such as `id: "codex"` requires that harness and fails closed if it is unavailable or fails.
* `id: "pi"` is accepted only as a deprecated alias for `openclaw` to preserve shipped configs from v2026.5.22 and earlier. New config should use `openclaw`.
* Runtime precedence is exact model policy first (`agents.list[].models["provider/model"]`, `agents.defaults.models["provider/model"]`, or `models.providers.<provider>.models[]`), then `agents.list[]` / `agents.defaults.models["provider/*"]`, then provider-wide policy at `models.providers.<provider>.agentRuntime`.
* Whole-agent runtime keys are legacy. `agents.defaults.agentRuntime`, `agents.list[].agentRuntime`, session runtime pins, and `OPENCLAW_AGENT_RUNTIME` are ignored by runtime selection. Run `openclaw doctor --fix` to remove stale values.
* OpenAI agent models use the Codex harness by default; provider/model `agentRuntime.id: "codex"` remains valid when you want to make that explicit.
* For Claude CLI deployments, prefer `model: "anthropic/claude-opus-4-8"` plus model-scoped `agentRuntime.id: "claude-cli"`. Legacy `claude-cli/<model>` refs still work for compatibility, but new config should keep provider/model selection canonical and put the execution backend in provider/model runtime policy.
* This only controls text agent-turn execution. Media generation, vision, PDF, music, video, and TTS still use their provider/model settings.

**Built-in alias shorthands** (only apply when the model is in `agents.defaults.models`):

| Alias               | Model                           |
| ------------------- | ------------------------------- |
| `opus`              | `anthropic/claude-opus-4-8`     |
| `sonnet`            | `anthropic/claude-sonnet-4-6`   |
| `gpt`               | `openai/gpt-5.4`                |
| `gpt-mini`          | `openai/gpt-5.4-mini`           |
| `gpt-nano`          | `openai/gpt-5.4-nano`           |
| `gemini`            | `google/gemini-3.1-pro-preview` |
| `gemini-flash`      | `google/gemini-3-flash-preview` |
| `gemini-flash-lite` | `google/gemini-3.1-flash-lite`  |

Your configured aliases always win over defaults.

Z.AI GLM-4.x models automatically enable thinking mode unless you set `--thinking off` or define `agents.defaults.models["zai/<model>"].params.thinking` yourself.
Z.AI models enable `tool_stream` by default for tool call streaming. Set `agents.defaults.models["zai/<model>"].params.tool_stream` to `false` to disable it.
Anthropic Claude Opus 4.8 keeps thinking off by default in OpenClaw; when adaptive thinking is explicitly enabled, Anthropic's provider-owned effort default is `high`. Claude 4.6 models default to `adaptive` when no explicit thinking level is set.

### `agents.defaults.cliBackends`

Optional CLI backends for text-only fallback runs (no tool calls). Useful as a backup when API providers fail.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      cliBackends: {
        "claude-cli": {
          command: "/opt/homebrew/bin/claude",
        },
        "my-cli": {
          command: "my-cli",
          args: ["--json"],
          output: "json",
          modelArg: "--model",
          sessionArg: "--session",
          sessionMode: "existing",
          systemPromptArg: "--system",
          // Or use systemPromptFileArg when the CLI accepts a prompt file flag.
          systemPromptWhen: "first",
          imageArg: "--image",
          imageMode: "repeat",
        },
      },
    },
  },
}
```

* CLI backends are text-first; tools are always disabled.
* Sessions supported when `sessionArg` is set.
* Image pass-through supported when `imageArg` accepts file paths.
* `reseedFromRawTranscriptWhenUncompacted: true` lets a backend recover safe
  invalidated sessions from a bounded raw OpenClaw transcript tail before the
  first compaction summary exists. Auth profile or credential-epoch changes
  still never raw-reseed.

### `agents.defaults.promptOverlays`

Provider-independent prompt overlays applied by model family on OpenClaw-assembled prompt surfaces. GPT-5-family model ids receive the shared behavior contract across OpenClaw/provider routes; `personality` controls only the friendly interaction-style layer. Native Codex app-server routes keep Codex-owned base/model instructions instead of this OpenClaw GPT-5 overlay, and OpenClaw disables Codex's built-in personality for native threads.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      promptOverlays: {
        gpt5: {
          personality: "friendly", // friendly | on | off
        },
      },
    },
  },
}
```

* `"friendly"` (default) and `"on"` enable the friendly interaction-style layer.
* `"off"` disables only the friendly layer; the tagged GPT-5 behavior contract remains enabled.
* Legacy `plugins.entries.openai.config.personality` is still read when this shared setting is unset.

### `agents.defaults.heartbeat`

Periodic heartbeat runs.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      heartbeat: {
        every: "30m", // 0m disables
        model: "openai/gpt-5.4-mini",
        includeReasoning: false,
        includeSystemPromptSection: true, // default: true; false omits the Heartbeat section from the system prompt
        lightContext: false, // default: false; true keeps only HEARTBEAT.md from workspace bootstrap files
        isolatedSession: false, // default: false; true runs each heartbeat in a fresh session (no conversation history)
        skipWhenBusy: false, // default: false; true also waits for this agent's subagent/nested lanes
        session: "main",
        to: "+15555550123",
        directPolicy: "allow", // allow (default) | block
        target: "none", // default: none | options: last | whatsapp | telegram | discord | ...
        prompt: "Read HEARTBEAT.md if it exists...",
        ackMaxChars: 300,
        suppressToolErrorWarnings: false,
        timeoutSeconds: 45,
      },
    },
  },
}
```

* `every`: duration string (ms/s/m/h). Default: `30m` (API-key auth) or `1h` (OAuth auth). Set to `0m` to disable.
* `includeSystemPromptSection`: when false, omits the Heartbeat section from the system prompt and skips `HEARTBEAT.md` injection into bootstrap context. Default: `true`.
* `suppressToolErrorWarnings`: when true, suppresses tool error warning payloads during heartbeat runs.
* `timeoutSeconds`: maximum time in seconds allowed for a heartbeat agent turn before it is aborted. Leave unset to use `agents.defaults.timeoutSeconds` when set, otherwise the heartbeat cadence capped at 600 seconds.
* `directPolicy`: direct/DM delivery policy. `allow` (default) permits direct-target delivery. `block` suppresses direct-target delivery and emits `reason=dm-blocked`.
* `lightContext`: when true, heartbeat runs use lightweight bootstrap context and keep only `HEARTBEAT.md` from workspace bootstrap files.
* `isolatedSession`: when true, each heartbeat runs in a fresh session with no prior conversation history. Same isolation pattern as cron `sessionTarget: "isolated"`. Reduces per-heartbeat token cost from \~100K to \~2-5K tokens.
* `skipWhenBusy`: when true, heartbeat runs defer on that agent's extra busy lanes: its own session-keyed subagent or nested command work. Cron lanes always defer heartbeats, even without this flag.
* Per-agent: set `agents.list[].heartbeat`. When any agent defines `heartbeat`, **only those agents** run heartbeats.
* Heartbeats run full agent turns — shorter intervals burn more tokens.

### `agents.defaults.compaction`

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      compaction: {
        mode: "safeguard", // default | safeguard
        provider: "my-provider", // id of a registered compaction provider plugin (optional)
        timeoutSeconds: 180,
        reserveTokensFloor: 24000,
        keepRecentTokens: 50000,
        recentTurnsPreserve: 3,
        maxHistoryShare: 0.7,
        identifierPolicy: "strict", // strict | off | custom
        identifierInstructions: "Preserve deployment IDs, ticket IDs, and host:port pairs exactly.", // used when identifierPolicy=custom
        qualityGuard: { enabled: true, maxRetries: 1 },
        midTurnPrecheck: { enabled: false }, // optional tool-loop pressure check
        postIndexSync: "async", // off | async | await
        postCompactionSections: ["Session Startup", "Red Lines"], // opt in to AGENTS.md section reinjection
        model: "openrouter/anthropic/claude-sonnet-4-6", // optional compaction-only model override
        truncateAfterCompaction: true, // rotate to a smaller successor JSONL after compaction
        maxActiveTranscriptBytes: "20mb", // optional preflight local compaction trigger
        notifyUser: true, // notices when compaction starts/completes and on memory-flush degradation (default: false)
        memoryFlush: {
          enabled: true,
          model: "ollama/qwen3:8b", // optional memory-flush-only model override
          softThresholdTokens: 6000,
          forceFlushTranscriptBytes: "2mb",
          systemPrompt: "Session nearing compaction. Store durable memories now.",
          prompt: "Write any lasting notes to memory/YYYY-MM-DD.md; reply with the exact silent token NO_REPLY if nothing to store.",
        },
      },
    },
  },
}
```

* `mode`: `default` or `safeguard` (chunked summarization for long histories). See [Compaction](/concepts/compaction).
* `provider`: id of a registered compaction provider plugin. When set, the provider's `summarize()` is called instead of built-in LLM summarization. Falls back to built-in on failure. Setting a provider forces `mode: "safeguard"`. See [Compaction](/concepts/compaction).
* `timeoutSeconds`: maximum seconds allowed for a single compaction operation before OpenClaw aborts it. Default: `180`.
* `reserveTokens`: token headroom kept available for model output and future tool results after compaction. When the model context window is known, OpenClaw caps the effective reserve so it cannot consume the prompt budget.
* `reserveTokensFloor`: minimum reserve enforced by the embedded runtime. Set `0` to disable the floor. The floor remains subject to the active context-window cap.
* `keepRecentTokens`: agent cut-point budget for keeping the most recent transcript tail verbatim. Manual `/compact` honors this when explicitly set; otherwise manual compaction is a hard checkpoint.
* `recentTurnsPreserve`: number of most recent user/assistant turns kept verbatim outside safeguard summarization. Default: `3`.
* `maxHistoryShare`: maximum fraction of the total context budget allowed for retained history after compaction (range `0.1`-`0.9`).
* `identifierPolicy`: `strict` (default), `off`, or `custom`. `strict` prepends built-in opaque identifier retention guidance during compaction summarization.
* `identifierInstructions`: optional custom identifier-preservation text used when `identifierPolicy=custom`.
* `qualityGuard`: retry-on-malformed-output checks for safeguard summaries. Enabled by default in safeguard mode; set `enabled: false` to skip the audit.
* `midTurnPrecheck`: optional tool-loop pressure check. When `enabled: true`, OpenClaw checks context pressure after tool results are appended and before the next model call. If the context no longer fits, it aborts the current attempt before submitting the prompt and reuses the existing precheck recovery path to truncate tool results or compact and retry. Works with both `default` and `safeguard` compaction modes. Default: disabled.
* `postIndexSync`: post-compaction session-memory reindex mode. Default: `"async"`. Use `"await"` for strongest freshness, `"async"` for lower compaction latency, or `"off"` only when session-memory sync is handled elsewhere.
* `postCompactionSections`: optional AGENTS.md H2/H3 section names to re-inject after compaction. Reinjection is disabled when unset or set to `[]`. Explicitly setting `["Session Startup", "Red Lines"]` enables that pair and preserves the legacy `Every Session`/`Safety` fallback. Enable this only when the extra context is worth the risk of duplicating project guidance already captured in the compaction summary.
* `model`: optional `provider/model-id` or bare alias from `agents.defaults.models` for compaction summarization only. Bare aliases resolve before dispatch; configured literal model IDs retain precedence on collisions. Use this when the main session should keep one model but compaction summaries should run on another; when unset, compaction uses the session's primary model.
* `truncateAfterCompaction`: rotates the active session JSONL after compaction so future turns load only the summary and unsummarized tail, while the previous full transcript remains archived. Prevents unbounded active transcript growth in long-running sessions. Default: `false`.
* `maxActiveTranscriptBytes`: optional byte threshold (`number` or strings like `"20mb"`) that triggers normal local compaction before a run when the active JSONL grows past the threshold. Requires `truncateAfterCompaction` so successful compaction can rotate to a smaller successor transcript. Disabled when unset or `0`.
* `notifyUser`: when `true`, sends brief context-maintenance notices to the user: when compaction starts and completes (for example, "Compacting context..." and "Compaction complete"), and when a pre-compaction memory flush is exhausted so the reply continues in a degraded state (for example, "Memory maintenance temporarily failed; continuing your reply."). Disabled by default to keep these notices silent.
* `memoryFlush`: silent agentic turn before auto-compaction to store durable memories. Set `model` to an exact provider/model such as `ollama/qwen3:8b` when this housekeeping turn should stay on a local model; the override does not inherit the active session fallback chain. `forceFlushTranscriptBytes` forces the flush when transcript file size reaches the threshold even if token counters are stale. Skipped when workspace is read-only.

### `agents.defaults.runRetries`

Outer run loop retry iteration boundaries for the embedded agent runtime to prevent infinite execution loops during failure recovery. This setting only applies to the embedded agent runtime, not ACP or CLI runtimes.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      runRetries: {
        base: 24,
        perProfile: 8,
        min: 32,
        max: 160,
      },
    },
    list: [
      {
        id: "main",
        runRetries: { max: 50 }, // optional per-agent overrides
      },
    ],
  },
}
```

* `base`: base number of run retry iterations for the outer run loop. Default: `24`.
* `perProfile`: additional run retry iterations granted per fallback profile candidate. Default: `8`.
* `min`: minimum absolute limit for run retry iterations. Default: `32`.
* `max`: maximum absolute limit for run retry iterations to prevent runaway execution. Default: `160`.

### `agents.defaults.contextPruning`

Prunes **old tool results** from in-memory context before sending to the LLM. Does **not** modify session history on disk. Disabled by default; set `mode: "cache-ttl"` to enable.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      contextPruning: {
        mode: "cache-ttl", // off (default) | cache-ttl
        ttl: "1h", // duration (ms/s/m/h), default unit: minutes; default: 5m
        keepLastAssistants: 3,
        softTrimRatio: 0.3,
        hardClearRatio: 0.5,
        minPrunableToolChars: 50000,
        softTrim: { maxChars: 4000, headChars: 1500, tailChars: 1500 },
        hardClear: { enabled: true, placeholder: "[Old tool result content cleared]" },
        tools: { deny: ["browser", "canvas"] },
      },
    },
  },
}
```

<Accordion title="cache-ttl mode behavior">
  * `mode: "cache-ttl"` enables pruning passes.
  * `ttl` controls how often pruning can run again (after the last cache touch). Default: `5m`.
  * Pruning soft-trims oversized tool results first, then hard-clears older tool results if needed.
  * `softTrimRatio` and `hardClearRatio` accept values from `0.0` through `1.0`; config validation rejects values outside that range.

  **Soft-trim** keeps beginning + end and inserts `...` in the middle.

  **Hard-clear** replaces the entire tool result with the placeholder.

  Notes:

  * Image blocks are never trimmed/cleared.
  * Ratios are character-based (approximate), not exact token counts.
  * If fewer than `keepLastAssistants` assistant messages exist, pruning is skipped.
</Accordion>

See [Session Pruning](/concepts/session-pruning) for behavior details.

### Block streaming

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      blockStreamingDefault: "off", // on | off
      blockStreamingBreak: "text_end", // text_end | message_end
      blockStreamingChunk: { minChars: 800, maxChars: 1200, breakPreference: "paragraph" },
      blockStreamingCoalesce: { idleMs: 1000 },
      humanDelay: { mode: "natural" }, // off (default) | natural | custom (use minMs/maxMs)
    },
  },
}
```

* Non-Telegram channels require explicit `*.blockStreaming: true` to enable block replies.
* Channel overrides: `channels.<channel>.blockStreamingCoalesce` (and per-account variants). Discord, Google Chat, Mattermost, MS Teams, Signal, and Slack default `minChars: 1500` / `idleMs: 1000`.
* `blockStreamingChunk.breakPreference`: preferred chunk boundary (`"paragraph" | "newline" | "sentence"`).
* `humanDelay`: randomized pause between block replies. Default: `off`. `natural` = 800-2500ms. `custom` uses `minMs`/`maxMs` (falls back to the natural range for any unset bound). Per-agent override: `agents.list[].humanDelay`.

See [Streaming](/concepts/streaming) for behavior + chunking details.

### Typing indicators

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      typingMode: "instant", // never | instant | thinking | message
      typingIntervalSeconds: 6,
    },
  },
}
```

* Defaults: `instant` for direct chats/mentions, `message` for unmentioned group chats.
* `typingIntervalSeconds` default: `6`.
* Per-session overrides: `session.typingMode`, `session.typingIntervalSeconds`.

See [Typing Indicators](/concepts/typing-indicators).

<a id="agentsdefaultssandbox" />

### `agents.defaults.sandbox`

Optional sandboxing for the embedded agent. See [Sandboxing](/gateway/sandboxing) for the full guide.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    defaults: {
      sandbox: {
        mode: "non-main", // off (default) | non-main | all
        backend: "docker", // docker (default) | ssh | openshell
        scope: "agent", // session | agent (default) | shared
        workspaceAccess: "none", // none (default) | ro | rw
        workspaceRoot: "~/.openclaw/sandboxes",
        docker: {
          image: "openclaw-sandbox:bookworm-slim",
          containerPrefix: "openclaw-sbx-",
          workdir: "/workspace",
          readOnlyRoot: true,
          tmpfs: ["/tmp", "/var/tmp", "/run"],
          network: "none",
          user: "1000:1000",
          capDrop: ["ALL"],
          env: { LANG: "C.UTF-8" },
          setupCommand: "apt-get update && apt-get install -y git curl jq",
          pidsLimit: 256,
          memory: "1g",
          memorySwap: "2g",
          cpus: 1,
          gpus: "all",
          ulimits: {
            nofile: { soft: 1024, hard: 2048 },
            nproc: 256,
          },
          seccompProfile: "/path/to/seccomp.json",
          apparmorProfile: "openclaw-sandbox",
          dns: ["1.1.1.1", "8.8.8.8"],
          extraHosts: ["internal.service:10.0.0.5"],
          binds: ["/home/user/source:/source:rw"],
        },
        ssh: {
          target: "user@gateway-host:22",
          command: "ssh",
          workspaceRoot: "/tmp/openclaw-sandboxes",
          strictHostKeyChecking: true,
          updateHostKeys: true,
          identityFile: "~/.ssh/id_ed25519",
          certificateFile: "~/.ssh/id_ed25519-cert.pub",
          knownHostsFile: "~/.ssh/known_hosts",
          // SecretRefs / inline contents also supported:
          // identityData: { source: "env", provider: "default", id: "SSH_IDENTITY" },
          // certificateData: { source: "env", provider: "default", id: "SSH_CERTIFICATE" },
          // knownHostsData: { source: "env", provider: "default", id: "SSH_KNOWN_HOSTS" },
        },
        browser: {
          enabled: false,
          image: "openclaw-sandbox-browser:bookworm-slim",
          network: "openclaw-sandbox-browser",
          cdpPort: 9222,
          cdpSourceRange: "172.21.0.1/32",
          vncPort: 5900,
          noVncPort: 6080,
          headless: false,
          enableNoVnc: true,
          allowHostControl: false,
          autoStart: true,
          autoStartTimeoutMs: 12000,
        },
        prune: {
          idleHours: 24,
          maxAgeDays: 7,
        },
      },
    },
  },
  tools: {
    sandbox: {
      tools: {
        allow: [
          "exec",
          "process",
          "read",
          "write",
          "edit",
          "apply_patch",
          "sessions_list",
          "sessions_history",
          "sessions_send",
          "sessions_spawn",
          "session_status",
        ],
        deny: ["browser", "canvas", "nodes", "cron", "discord", "gateway"],
      },
    },
  },
}
```

Defaults shown above (`off`/`docker`/`agent`/`none`/`bookworm-slim` image/`none` network/etc.) are the actual OpenClaw defaults, not just illustrative values.

<Accordion title="Sandbox details">
  **Backend:**

  * `docker`: local Docker runtime (default)
  * `ssh`: generic SSH-backed remote runtime
  * `openshell`: OpenShell runtime

  When `backend: "openshell"` is selected, runtime-specific settings move to
  `plugins.entries.openshell.config`.

  **SSH backend config:**

  * `target`: SSH target in `user@host[:port]` form
  * `command`: SSH client command (default: `ssh`)
  * `workspaceRoot`: absolute remote root used for per-scope workspaces (default: `/tmp/openclaw-sandboxes`)
  * `identityFile` / `certificateFile` / `knownHostsFile`: existing local files passed to OpenSSH
  * `identityData` / `certificateData` / `knownHostsData`: inline contents or SecretRefs that OpenClaw materializes into temp files at runtime
  * `strictHostKeyChecking` / `updateHostKeys`: OpenSSH host-key policy knobs (both default `true`)

  **SSH auth precedence:**

  * `identityData` wins over `identityFile`
  * `certificateData` wins over `certificateFile`
  * `knownHostsData` wins over `knownHostsFile`
  * SecretRef-backed `*Data` values are resolved from the active secrets runtime snapshot before the sandbox session starts

  **SSH backend behavior:**

  * seeds the remote workspace once after create or recreate
  * then keeps the remote SSH workspace canonical
  * routes `exec`, file tools, and media paths over SSH
  * does not sync remote changes back to the host automatically
  * does not support sandbox browser containers

  **Workspace access:**

  * `none`: per-scope sandbox workspace under `~/.openclaw/sandboxes` (default)
  * `ro`: sandbox workspace at `/workspace`, agent workspace mounted read-only at `/agent`
  * `rw`: agent workspace mounted read/write at `/workspace`

  **Scope:**

  * `session`: per-session container + workspace
  * `agent`: one container + workspace per agent (default)
  * `shared`: shared container and workspace (no cross-session isolation)

  **OpenShell plugin config:**

  ```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
  {
    plugins: {
      entries: {
        openshell: {
          enabled: true,
          config: {
            mode: "mirror", // mirror (default) | remote
            command: "openshell",
            from: "openclaw",
            remoteWorkspaceDir: "/sandbox",
            remoteAgentWorkspaceDir: "/agent",
            gateway: "lab", // optional
            gatewayEndpoint: "https://lab.example", // optional
            policy: "strict", // optional OpenShell policy id
            providers: ["openai"], // optional
            autoProviders: true,
            timeoutSeconds: 120,
          },
        },
      },
    },
  }
  ```

  **OpenShell mode:**

  * `mirror`: seed remote from local before exec, sync back after exec; local workspace stays canonical
  * `remote`: seed remote once when the sandbox is created, then keep the remote workspace canonical

  In `remote` mode, host-local edits made outside OpenClaw are not synced into the sandbox automatically after the seed step.
  Transport is SSH into the OpenShell sandbox, but the plugin owns sandbox lifecycle and optional mirror sync.

  **`setupCommand`** runs once after container creation (via `sh -lc`). Needs network egress, writable root, root user.

  **Containers default to `network: "none"`** — set to `"bridge"` (or a custom bridge network) if the agent needs outbound access.
  `"host"` is blocked. `"container:<id>"` is blocked by default unless you explicitly set
  `sandbox.docker.dangerouslyAllowContainerNamespaceJoin: true` (break-glass).
  Codex app-server turns in an active OpenClaw sandbox use this same egress setting for their native code-mode network access.

  **Inbound attachments** are staged into `media/inbound/*` in the active workspace.

  **`docker.binds`** mounts additional host directories; global and per-agent binds are merged.

  **Sandboxed browser** (`sandbox.browser.enabled`, default `false`): Chromium + CDP in a container. noVNC URL injected into system prompt. Does not require `browser.enabled` in `openclaw.json`.
  noVNC observer access uses VNC auth by default and OpenClaw emits a short-lived token URL (instead of exposing the password in the shared URL).

  * `allowHostControl: false` (default) blocks sandboxed sessions from targeting the host browser.
  * `network` defaults to `openclaw-sandbox-browser` (dedicated bridge network). Set to `bridge` only when you explicitly want global bridge connectivity. `"host"` is blocked here too.
  * `cdpSourceRange` optionally restricts CDP ingress at the container edge to a CIDR range (for example `172.21.0.1/32`).
  * `sandbox.browser.binds` mounts additional host directories into the sandbox browser container only. When set (including `[]`), it replaces `docker.binds` for the browser container.
  * The sandbox browser container's Chromium always launches with `--no-sandbox --disable-setuid-sandbox` (containers do not have the kernel primitives Chrome's own sandbox needs); there is no config toggle for this.
  * Launch defaults are defined in `scripts/sandbox-browser-entrypoint.sh` and tuned for container hosts:
    * `--remote-debugging-address=127.0.0.1`
    * `--remote-debugging-port=<derived from OPENCLAW_BROWSER_CDP_PORT>`
    * `--user-data-dir=${HOME}/.chrome`
    * `--no-first-run`
    * `--no-default-browser-check`
    * `--disable-dev-shm-usage`
    * `--disable-background-networking`
    * `--disable-breakpad`
    * `--disable-crash-reporter`
    * `--no-zygote`
    * `--metrics-recording-only`
    * `--password-store=basic`
    * `--use-mock-keychain`
    * `--disable-3d-apis`, `--disable-gpu`, and `--disable-software-rasterizer` are
      enabled by default and can be disabled with
      `OPENCLAW_BROWSER_DISABLE_GRAPHICS_FLAGS=0` if WebGL/3D usage requires it.
    * `--disable-extensions` (default enabled); `OPENCLAW_BROWSER_DISABLE_EXTENSIONS=0`
      re-enables extensions if your workflow depends on them.
    * `--renderer-process-limit=2` by default; change with
      `OPENCLAW_BROWSER_RENDERER_PROCESS_LIMIT=<N>`, set `0` to use Chromium's
      default process limit.
    * `--headless=new` only when `headless` is enabled.
    * Defaults are the container image baseline; use a custom browser image with a custom
      entrypoint to change container defaults.
</Accordion>

Browser sandboxing and `sandbox.docker.binds` are Docker-only.

Build images (from a source checkout):

```bash theme={"theme":{"light":"min-light","dark":"min-dark"}}
scripts/sandbox-setup.sh           # main sandbox image
scripts/sandbox-browser-setup.sh   # optional browser image
```

For npm installs without a source checkout, see [Sandboxing § Images and setup](/gateway/sandboxing#images-and-setup) for inline `docker build` commands.

### `agents.list` (per-agent overrides)

Use `agents.list[].tts` to give an agent its own TTS provider, voice, model,
style, or auto-TTS mode. The agent block deep-merges over global
`messages.tts`, so shared credentials can stay in one place while individual
agents override only the voice or provider fields they need. The active agent's
override applies to automatic spoken replies, `/tts audio`, `/tts status`, and
the `tts` agent tool. See [Text-to-speech](/tools/tts#per-agent-voice-overrides)
for provider examples and precedence.

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    list: [
      {
        id: "main",
        default: true,
        name: "Main Agent",
        workspace: "~/.openclaw/workspace",
        agentDir: "~/.openclaw/agents/main/agent",
        model: "anthropic/claude-opus-4-6", // or { primary, fallbacks }
        utilityModel: "openai/gpt-5.4-mini",
        thinkingDefault: "high", // per-agent thinking level override
        reasoningDefault: "on", // per-agent reasoning visibility override
        fastModeDefault: false, // per-agent fast mode override
        params: { cacheRetention: "none" }, // overrides matching defaults.models params by key
        tts: {
          providers: {
            elevenlabs: { speakerVoiceId: "EXAVITQu4vr4xnSDxMaL" },
          },
        },
        skills: ["docs-search"], // replaces agents.defaults.skills when set
        identity: {
          name: "Samantha",
          theme: "helpful sloth",
          emoji: "🦥",
          avatar: "avatars/samantha.png",
        },
        groupChat: { mentionPatterns: ["@openclaw"] },
        sandbox: { mode: "off" },
        runtime: {
          type: "acp",
          acp: {
            agent: "codex",
            backend: "acpx",
            mode: "persistent", // persistent | oneshot
            cwd: "/workspace/openclaw",
          },
        },
        subagents: { allowAgents: ["*"] },
        tools: {
          profile: "coding",
          allow: ["browser"],
          deny: ["canvas"],
          elevated: { enabled: true },
        },
      },
    ],
  },
}
```

* `id`: stable agent id (required).
* `default`: when multiple are set, first wins (warning logged). If none set, first list entry is default.
* `model`: string form sets a strict per-agent primary with no model fallback; object form `{ primary }` is also strict unless you add `fallbacks`. Use `{ primary, fallbacks: [...] }` to opt that agent into fallback, or `{ primary, fallbacks: [] }` to make strict behavior explicit. Cron jobs that only override `primary` still inherit default fallbacks unless you set `fallbacks: []`.
* `utilityModel`: optional per-agent override for short internal tasks such as generated session and thread titles. Falls back to `agents.defaults.utilityModel`, then this agent's primary model.
* `params`: per-agent stream params merged over the selected model entry in `agents.defaults.models`. Use this for agent-specific overrides like `cacheRetention`, `temperature`, or `maxTokens` without duplicating the whole model catalog.
* `tts`: optional per-agent text-to-speech overrides. The block deep-merges over `messages.tts`, so keep shared provider credentials and fallback policy in `messages.tts` and set only persona-specific values such as provider, voice, model, style, or auto mode here.
* `skills`: optional per-agent skill allowlist. If omitted, the agent inherits `agents.defaults.skills` when set; an explicit list replaces defaults instead of merging, and `[]` means no skills.
* `thinkingDefault`: optional per-agent default thinking level (`off | minimal | low | medium | high | xhigh | adaptive | max`). Overrides `agents.defaults.thinkingDefault` for this agent when no per-message or session override is set. The selected provider/model profile controls which values are valid; for Google Gemini, `adaptive` keeps provider-owned dynamic thinking (`thinkingLevel` omitted on Gemini 3/3.1, `thinkingBudget: -1` on Gemini 2.5).
* `reasoningDefault`: optional per-agent default reasoning visibility (`on | off | stream`). Overrides `agents.defaults.reasoningDefault` for this agent when no per-message or session reasoning override is set.
* `fastModeDefault`: optional per-agent default for fast mode (`"auto" | true | false`). Applies when no per-message or session fast-mode override is set.
* `models`: optional per-agent model catalog/runtime overrides keyed by full `provider/model` ids. Use `models["provider/model"].agentRuntime` for per-agent runtime exceptions.
* `runtime`: optional per-agent runtime descriptor. Use `type: "acp"` with `runtime.acp` defaults (`agent`, `backend`, `mode`, `cwd`) when the agent should default to ACP harness sessions.
* `identity.avatar`: workspace-relative path, `http(s)` URL, or `data:` URI.
* Local workspace-relative `identity.avatar` image files are limited to 2 MB. `http(s)` URLs and `data:` URIs are not checked against the local file-size limit.
* `identity` derives defaults: `ackReaction` from `emoji`, `mentionPatterns` from `name`/`emoji`.
* `subagents.allowAgents`: allowlist of configured agent ids for explicit `sessions_spawn.agentId` targets (`["*"]` = any configured target; default: same agent only). Include the requester id when self-targeted `agentId` calls should be allowed. Stale entries whose agent config was deleted are rejected by `sessions_spawn` and omitted from `agents_list`; run `openclaw doctor --fix` to clean them up, or add a minimal `agents.list[]` entry if that target should remain spawnable while inheriting defaults.
* Sandbox inheritance guard: if the requester session is sandboxed, `sessions_spawn` rejects targets that would run unsandboxed.
* `subagents.requireAgentId`: when true, block `sessions_spawn` calls that omit `agentId` (forces explicit profile selection; default: false).
* `subagents.maxConcurrent`: max concurrent child-agent runs across subagent execution. Default: `8`.
* `subagents.maxChildrenPerAgent`: max active children a single agent session can spawn. Default: `5`.
* `subagents.maxSpawnDepth`: max nesting depth for sub-agent spawning (`1`-`5`). Default: `1` (no nesting).
* `subagents.archiveAfterMinutes`: age before completed subagent state is archived. Default: `60`.

***

## Multi-agent routing

Run multiple isolated agents inside one Gateway. See [Multi-Agent](/concepts/multi-agent).

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  agents: {
    list: [
      { id: "home", default: true, workspace: "~/.openclaw/workspace-home" },
      { id: "work", workspace: "~/.openclaw/workspace-work" },
    ],
  },
  bindings: [
    { agentId: "home", match: { channel: "whatsapp", accountId: "personal" } },
    { agentId: "work", match: { channel: "whatsapp", accountId: "biz" } },
  ],
}
```

### Binding match fields

* `type` (optional): `route` for normal routing (missing type defaults to route), `acp` for persistent ACP conversation bindings.
* `match.channel` (required)
* `match.accountId` (optional; `*` = any account; omitted = default account)
* `match.peer` (optional; `{ kind: direct|group|channel, id }`)
* `match.guildId` / `match.teamId` (optional; channel-specific)
* `acp` (optional; only for `type: "acp"`): `{ mode, label, cwd, backend }`

**Deterministic match order:**

1. `match.peer`
2. `match.guildId`
3. `match.teamId`
4. `match.accountId` (exact, no peer/guild/team)
5. `match.accountId: "*"` (channel-wide)
6. Default agent

Within each tier, the first matching `bindings` entry wins.

For `type: "acp"` entries, OpenClaw resolves by exact conversation identity (`match.channel` + account + `match.peer.id`) and does not use the route binding tier order above.

### Per-agent access profiles

<Accordion title="Full access (no sandbox)">
  ```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
  {
    agents: {
      list: [
        {
          id: "personal",
          workspace: "~/.openclaw/workspace-personal",
          sandbox: { mode: "off" },
        },
      ],
    },
  }
  ```
</Accordion>

<Accordion title="Read-only tools + workspace">
  ```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
  {
    agents: {
      list: [
        {
          id: "family",
          workspace: "~/.openclaw/workspace-family",
          sandbox: { mode: "all", scope: "agent", workspaceAccess: "ro" },
          tools: {
            allow: [
              "read",
              "sessions_list",
              "sessions_history",
              "sessions_send",
              "sessions_spawn",
              "session_status",
            ],
            deny: ["write", "edit", "apply_patch", "exec", "process", "browser"],
          },
        },
      ],
    },
  }
  ```
</Accordion>

<Accordion title="No filesystem access (messaging only)">
  ```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
  {
    agents: {
      list: [
        {
          id: "public",
          workspace: "~/.openclaw/workspace-public",
          sandbox: { mode: "all", scope: "agent", workspaceAccess: "none" },
          tools: {
            allow: [
              "sessions_list",
              "sessions_history",
              "sessions_send",
              "sessions_spawn",
              "session_status",
              "whatsapp",
              "telegram",
              "slack",
              "discord",
              "gateway",
            ],
            deny: [
              "read",
              "write",
              "edit",
              "apply_patch",
              "exec",
              "process",
              "browser",
              "canvas",
              "nodes",
              "cron",
              "gateway",
              "image",
            ],
          },
        },
      ],
    },
  }
  ```
</Accordion>

See [Multi-Agent Sandbox & Tools](/tools/multi-agent-sandbox-tools) for precedence details.

***

## Session

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  session: {
    scope: "per-sender",
    dmScope: "main", // main | per-peer | per-channel-peer | per-account-channel-peer
    identityLinks: {
      alice: ["telegram:123456789", "discord:987654321012345678"],
    },
    reset: {
      mode: "daily", // daily | idle
      atHour: 4,
      idleMinutes: 60,
    },
    resetByType: {
      thread: { mode: "daily", atHour: 4 },
      direct: { mode: "idle", idleMinutes: 240 },
      group: { mode: "idle", idleMinutes: 120 },
    },
    resetByChannel: {
      discord: { mode: "idle", idleMinutes: 30 },
    },
    resetTriggers: ["/new", "/reset"],
    store: "~/.openclaw/agents/{agentId}/sessions/sessions.json",
    maintenance: {
      mode: "enforce", // enforce (default) | warn
      pruneAfter: "30d",
      maxEntries: 500,
      resetArchiveRetention: "30d", // duration or false
      maxDiskBytes: "500mb", // optional hard budget
      highWaterBytes: "400mb", // optional cleanup target
    },
    writeLock: {
      acquireTimeoutMs: 60000,
      staleMs: 1800000,
      maxHoldMs: 300000,
    },
    threadBindings: {
      enabled: true,
      idleHours: 24, // default inactivity auto-unfocus in hours (`0` disables)
      maxAgeHours: 0, // default hard max age in hours (`0` disables)
    },
    mainKey: "main", // legacy (runtime always uses "main")
    agentToAgent: { maxPingPongTurns: 5 },
    sendPolicy: {
      rules: [{ action: "deny", match: { channel: "discord", chatType: "group" } }],
      default: "allow",
    },
  },
}
```

<Accordion title="Session field details">
  * **`scope`**: base session grouping strategy for group-chat contexts.
    * `per-sender` (default): each sender gets an isolated session within a channel context.
    * `global`: all participants in a channel context share a single session (use only when shared context is intended).
  * **`dmScope`**: how DMs are grouped.
    * `main`: all DMs share the main session.
    * `per-peer`: isolate by sender id across channels.
    * `per-channel-peer`: isolate per channel + sender (recommended for multi-user inboxes).
    * `per-account-channel-peer`: isolate per account + channel + sender (recommended for multi-account).
  * **`identityLinks`**: map canonical ids to provider-prefixed peers for cross-channel session sharing. Dock commands such as `/dock_discord` use the same map to switch the active session's reply route to another linked channel peer; see [Channel docking](/concepts/channel-docking).
  * **`reset`**: primary reset policy. `daily` resets at `atHour` local time; `idle` resets after `idleMinutes`. When both configured, whichever expires first wins. Daily reset freshness uses the session row's `sessionStartedAt`; idle reset freshness uses `lastInteractionAt`. Background/system-event writes such as heartbeat, cron wakeups, exec notifications, and gateway bookkeeping can update `updatedAt`, but they do not keep daily/idle sessions fresh.
  * **`resetByType`**: per-type overrides (`direct`, `group`, `thread`). Legacy `dm` accepted as alias for `direct`.
  * **`resetByChannel`**: per-channel reset overrides keyed by provider/channel id. When the session's channel has a matching entry, it wins outright over `resetByType`/`reset` for that session. Use only when one channel needs reset behavior different from the type-level policy.
  * **`mainKey`**: legacy field. Runtime always uses `"main"` for the main direct-chat bucket.
  * **`agentToAgent.maxPingPongTurns`**: maximum reply-back turns between agents during agent-to-agent exchanges (integer, range: `0`-`20`, default: `5`). `0` disables ping-pong chaining.
  * **`sendPolicy`**: match by `channel`, `chatType` (`direct|group|channel`, with legacy `dm` alias), `keyPrefix`, or `rawKeyPrefix`. First deny wins.
  * **`maintenance`**: session-store cleanup + retention controls.
    * `mode`: `enforce` applies cleanup and is the default; `warn` emits warnings only.
    * `pruneAfter`: age cutoff for stale entries (default `30d`).
    * `maxEntries`: maximum number of entries in `sessions.json` (default `500`). Runtime writes batch cleanup with a small high-water buffer for production-sized caps; `openclaw sessions cleanup --enforce` applies the cap immediately.
    * Short-lived gateway model-run probe sessions use fixed `24h` retention, but cleanup is pressure-gated: it only removes stale strict model-run probe rows when session-entry maintenance/cap pressure is reached. Only strict explicit probe keys matching `agent:*:explicit:model-run-<uuid>` are eligible; normal direct, group, thread, cron, hook, heartbeat, ACP, and sub-agent sessions do not inherit this 24h retention. When model-run cleanup runs, it runs before the broader `pruneAfter` stale-entry cleanup and `maxEntries` cap.
    * `rotateBytes`: deprecated and ignored; `openclaw doctor --fix` removes it from older configs.
    * `resetArchiveRetention`: retention for `*.reset.<timestamp>` transcript archives. Defaults to `pruneAfter`; set `false` to disable.
    * `maxDiskBytes`: optional sessions-directory disk budget. In `warn` mode it logs warnings; in `enforce` mode it removes oldest artifacts/sessions first.
    * `highWaterBytes`: optional target after budget cleanup. Defaults to `80%` of `maxDiskBytes`.
  * **`writeLock`**: session transcript write-lock controls. Tune only when legitimate transcript prep, cleanup, compaction, or mirror work contends longer than the default policies.
    * `acquireTimeoutMs`: milliseconds to wait while acquiring a lock before reporting the session as busy. Default: `60000`; env override `OPENCLAW_SESSION_WRITE_LOCK_ACQUIRE_TIMEOUT_MS`.
    * `staleMs`: milliseconds before an existing lock is treated as stale and reclaimed. Default: `1800000`; env override `OPENCLAW_SESSION_WRITE_LOCK_STALE_MS`.
    * `maxHoldMs`: milliseconds a held in-process lock may remain held before the watchdog releases it. Default: `300000`; env override `OPENCLAW_SESSION_WRITE_LOCK_MAX_HOLD_MS`.
  * **`threadBindings`**: global defaults for thread-bound session features.
    * `enabled`: master default switch (providers can override; Discord uses `channels.discord.threadBindings.enabled`)
    * `idleHours`: default inactivity auto-unfocus in hours (`0` disables; providers can override)
    * `maxAgeHours`: default hard max age in hours (`0` disables; providers can override)
    * `spawnSessions`: default gate for creating thread-bound work sessions from `sessions_spawn` and ACP thread spawns. Defaults to `true` when thread bindings are enabled; providers/accounts can override.
    * `defaultSpawnContext`: default native subagent context for thread-bound spawns (`"fork"` or `"isolated"`). Defaults to `"fork"`.
</Accordion>

***

## Messages

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  messages: {
    responsePrefix: "🦞", // or "auto"
    ackReaction: "👀",
    ackReactionScope: "group-mentions", // group-mentions | group-all | direct | all | off | none
    removeAckAfterReply: false,
    queue: {
      mode: "steer", // steer (default) | followup | collect | interrupt
      debounceMs: 500,
      cap: 20,
      drop: "summarize", // old | new | summarize (default)
      byChannel: {
        whatsapp: "followup",
        telegram: "followup",
      },
    },
    inbound: {
      debounceMs: 2000, // 0 disables
      byChannel: {
        whatsapp: 5000,
        slack: 1500,
      },
    },
  },
}
```

### Response prefix

Per-channel/account overrides: `channels.<channel>.responsePrefix`, `channels.<channel>.accounts.<id>.responsePrefix`.

Resolution (most specific wins): account → channel → global. `""` disables and stops cascade. `"auto"` derives `[{identity.name}]`.

**Template variables:**

| Variable          | Description            | Example                     |
| ----------------- | ---------------------- | --------------------------- |
| `{model}`         | Short model name       | `claude-opus-4-6`           |
| `{modelFull}`     | Full model identifier  | `anthropic/claude-opus-4-6` |
| `{provider}`      | Provider name          | `anthropic`                 |
| `{thinkingLevel}` | Current thinking level | `high`, `low`, `off`        |
| `{identity.name}` | Agent identity name    | (same as `"auto"`)          |

Variables are case-insensitive. `{think}` is an alias for `{thinkingLevel}`.

### Ack reaction

* Defaults to active agent's `identity.emoji`, otherwise `"👀"`. Set `""` to disable.
* Per-channel overrides: `channels.<channel>.ackReaction`, `channels.<channel>.accounts.<id>.ackReaction`.
* Resolution order: account → channel → `messages.ackReaction` → identity fallback.
* Scope: `group-mentions` (default), `group-all`, `direct`, `all`, or `off`/`none` (disables ack reactions entirely).
* `removeAckAfterReply`: removes ack after reply on reaction-capable channels such as Slack, Discord, Signal, Telegram, WhatsApp, and iMessage.
* `messages.statusReactions.enabled`: enables lifecycle status reactions on Slack, Discord, Signal, Telegram, and WhatsApp.
  On Discord, unset keeps status reactions enabled when ack reactions are active.
  On Slack, Signal, Telegram, and WhatsApp, set it explicitly to `true` to enable lifecycle status reactions.
  Slack uses its native assistant thread status and rotating loading messages for progress by default, while keeping the configured ack reaction static.
* `messages.statusReactions.emojis`: overrides lifecycle emoji keys:
  `queued`, `thinking`, `compacting`, `tool`, `coding`, `web`, `deploy`, `build`,
  `concierge`, `done`, `error`, `stallSoft`, and `stallHard`.
  Telegram only allows a fixed reaction set, so unsupported configured emoji fall back
  to the nearest supported status variant for that chat.

### Queue

* `mode`: queue strategy for inbound messages that arrive while a session run is active. Default: `"steer"`.
  * `steer`: inject the new prompt into the active run.
  * `followup`: run the new prompt after the active run finishes.
  * `collect`: batch compatible messages and run them together later.
  * `interrupt`: abort the active run before starting the newest prompt.
* `debounceMs`: delay before dispatching a queued/steered message. Default: `500`.
* `cap`: maximum queued messages before the drop policy applies. Default: `20`.
* `drop`: strategy when the cap is exceeded. `"summarize"` (default) drops oldest entries but keeps compact summaries; `"old"` drops oldest without summaries; `"new"` rejects the newest item.
* `byChannel`: per-channel `mode` overrides keyed by provider id.
* `debounceMsByChannel`: per-channel `debounceMs` overrides keyed by provider id.

### Inbound debounce

Batches rapid text-only messages from the same sender into a single agent turn. Media/attachments flush immediately. Control commands bypass debouncing. Default `debounceMs`: `2000`.

### Other message keys

* `messages.messagePrefix`: prefix text prepended to inbound user messages before they reach the agent runtime. Use sparingly for channel context markers.
* `messages.visibleReplies`: controls visible source replies across direct, group, and channel conversations (`"message_tool"` requires `message(action=send)` for visible output; `"automatic"` posts normal replies as before).
* `messages.usageTemplate` / `messages.responseUsage`: custom `/usage` footer template and default per-reply usage mode (`off | tokens | full`, plus legacy `on` alias for `tokens`).
* `messages.groupChat.mentionPatterns` / `historyLimit`: group-message mention triggers and history window sizing.
* `messages.suppressToolErrors`: when `true`, suppresses `⚠️` tool-error warnings shown to the user (the agent still sees errors in context and can retry). Default: `false`.

### TTS (text-to-speech)

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  messages: {
    tts: {
      auto: "off", // off (default) | always | inbound | tagged
      mode: "final", // final | all
      provider: "elevenlabs",
      summaryModel: "openai/gpt-5.4-mini",
      modelOverrides: { enabled: true },
      maxTextLength: 4000,
      timeoutMs: 30000,
      prefsPath: "~/.openclaw/settings/tts.json",
      providers: {
        elevenlabs: {
          apiKey: "elevenlabs_api_key",
          baseUrl: "https://api.elevenlabs.io",
          speakerVoiceId: "voice_id",
          modelId: "eleven_multilingual_v2",
          seed: 42,
          applyTextNormalization: "auto",
          languageCode: "en",
          voiceSettings: {
            stability: 0.5,
            similarityBoost: 0.75,
            style: 0.0,
            useSpeakerBoost: true,
            speed: 1.0,
          },
        },
        microsoft: {
          speakerVoice: "en-US-MichelleNeural",
          lang: "en-US",
          outputFormat: "audio-24khz-48kbitrate-mono-mp3",
        },
        openai: {
          apiKey: "openai_api_key",
          baseUrl: "https://api.openai.com/v1",
          model: "gpt-4o-mini-tts",
          speakerVoice: "coral",
        },
      },
    },
  },
}
```

* `auto` controls the default auto-TTS mode: `off`, `always`, `inbound`, or `tagged`. `/tts on|off` can override local prefs, and `/tts status` shows the effective state.
* `summaryModel` overrides `agents.defaults.model.primary` for auto-summary.
* `modelOverrides` is enabled by default (`enabled !== false`); `modelOverrides.allowProvider` is opt-in.
* API keys fall back to `ELEVENLABS_API_KEY`/`XI_API_KEY` and `OPENAI_API_KEY`.
* Bundled speech providers are plugin-owned. If `plugins.allow` is set, include each TTS provider plugin you want to use, for example `microsoft` for Edge TTS. The legacy `edge` provider id is accepted as an alias for `microsoft`.
* `providers.openai.baseUrl` overrides the OpenAI TTS endpoint. Resolution order is config, then `OPENAI_TTS_BASE_URL`, then `https://api.openai.com/v1`.
* When `providers.openai.baseUrl` points to a non-OpenAI endpoint, OpenClaw treats it as an OpenAI-compatible TTS server and relaxes model/voice validation.

***

## Talk

Defaults for Talk mode (macOS/iOS/Android and the browser Control UI).

```json5 theme={"theme":{"light":"min-light","dark":"min-dark"}}
{
  talk: {
    provider: "elevenlabs",
    providers: {
      elevenlabs: {
        speakerVoiceId: "elevenlabs_voice_id",
        voiceAliases: {
          Clawd: "EXAVITQu4vr4xnSDxMaL",
          Roger: "CwhRBWXzGAHq8TQ4Fs17",
        },
        modelId: "eleven_multilingual_v2",
        outputFormat: "mp3_44100_128",
        apiKey: "elevenlabs_api_key",
      },
      mlx: {
        modelId: "mlx-community/Soprano-80M-bf16",
      },
      system: {},
    },
    consultThinkingLevel: "low",
    consultFastMode: true,
    speechLocale: "ru-RU",
    silenceTimeoutMs: 1500,
    interruptOnSpeech: true,
    realtime: {
      provider: "openai",
      providers: {
        openai: {
          model: "gpt-realtime-2",
          speakerVoice: "cedar",
        },
      },
      instructions: "Speak warmly and keep answers brief.",
      mode: "realtime", // realtime | stt-tts | transcription
      transport: "webrtc", // webrtc | provider-websocket | gateway-relay | managed-room
      vadThreshold: 0.5,
      silenceDurationMs: 500,
      prefixPaddingMs: 300,
      reasoningEffort: "medium",
      brain: "agent-consult", // agent-consult | direct-tools | none
    },
  },
}
```

* `talk.provider` must match a key in `talk.providers` when multiple Talk providers are configured.
* Legacy flat Talk keys (`talk.voiceId`, `talk.voiceAliases`, `talk.modelId`, `talk.outputFormat`, `talk.apiKey`) are compatibility-only. Run `openclaw doctor --fix` to rewrite persisted config into `talk.providers.<provider>`.
* Voice IDs fall back to `ELEVENLABS_VOICE_ID` or `SAG_VOICE_ID` (macOS Talk client behavior).
* `providers.*.apiKey` accepts plaintext strings or SecretRef objects.
* `ELEVENLABS_API_KEY` fallback applies only when no Talk API key is configured.
* `providers.*.voiceAliases` lets Talk directives use friendly names.
* `providers.mlx.modelId` selects the Hugging Face repo used by the macOS local MLX helper. If omitted, macOS uses `mlx-community/Soprano-80M-bf16`.
* macOS MLX playback runs through the bundled `openclaw-mlx-tts` helper when present, or an executable on `PATH`; `OPENCLAW_MLX_TTS_BIN` overrides the helper path for development.
* `consultThinkingLevel` controls the thinking level for the full OpenClaw agent run behind Control UI Talk realtime `openclaw_agent_consult` calls. Leave unset to preserve normal session/model behavior.
* `consultFastMode` sets a one-shot fast-mode override for Control UI Talk realtime consults without changing the session's normal fast-mode setting.
* `speechLocale` sets the BCP 47 locale id used by iOS/macOS Talk speech recognition. Leave unset to use the device default.
* `silenceTimeoutMs` controls how long Talk mode waits after user silence before it sends the transcript. Unset keeps the platform default pause window (`700 ms on macOS and Android, 900 ms on iOS`).
* `realtime.instructions` appends provider-facing system instructions to OpenClaw's built-in realtime prompt, so voice style can be configured without losing default `openclaw_agent_consult` guidance.
* `realtime.vadThreshold` sets the provider voice-activity threshold from `0` (most sensitive) to `1` (least sensitive). Unset keeps the provider default.
* `realtime.silenceDurationMs` sets the positive whole-number silence window before the provider commits a realtime user turn. Unset keeps the provider default.
* `realtime.prefixPaddingMs` sets the non-negative whole-number amount of audio retained before detected speech begins. Unset keeps the provider default.
* `realtime.reasoningEffort` sets the provider-specific reasoning level for realtime sessions. Unset keeps the provider default.
* `realtime.consultRouting`: `"provider-direct"` (default) preserves direct provider replies when the realtime provider produces a final user transcript without `openclaw_agent_consult`. `"force-agent-consult"` routes the finalized request through OpenClaw instead.

***

## Related

* [Configuration reference](/gateway/configuration-reference) — all other config keys
* [Configuration](/gateway/configuration) — common tasks and quick setup
* [Configuration examples](/gateway/configuration-examples)