Tools and plugins

Everything the agent does beyond generating text happens through tools. Tools are how the agent reads files, runs commands, browses the web, sends messages, and interacts with devices.

Tools, skills, and plugins

OpenClaw has three layers that work together:

The agent sees tools as structured function definitions sent to the model API.

[Skills reference](/tools/skills) | [Creating skills](/tools/creating-skills)

[Install and configure plugins](/tools/plugin) | [Build your own](/plugins/building-plugins)

Built-in tools

These tools ship with OpenClaw and are available without installing any plugins:

For image work, use image for analysis and image_generate for generation or editing. If you target openai/*, google/*, fal/*, or another non-default image provider, configure that provider's auth/API key first.

For music work, use music_generate. If you target google/*, minimax/*, or another non-default music provider, configure that provider's auth/API key first.

For video work, use video_generate. If you target qwen/* or another non-default video provider, configure that provider's auth/API key first.

For workflow-driven audio generation, use music_generate when a plugin such as ComfyUI registers it. This is separate from tts, which is text-to-speech.

session_status is the lightweight status/readback tool in the sessions group. It answers /status-style questions about the current session and can optionally set a per-session model override; model=default clears that override. Like /status, it can backfill sparse token/cache counters and the active runtime model label from the latest transcript usage entry.

gateway is the owner-only runtime tool for gateway operations:

• config.schema.lookup for one path-scoped config subtree before edits
• config.get for the current config snapshot + hash
• config.patch for partial config updates with restart
• config.apply only for full-config replacement
• update.run for explicit self-update + restart

For partial changes, prefer config.schema.lookup then config.patch. Use config.apply only when you intentionally replace the entire config. For broader config docs, read Configuration and Configuration reference. The tool also refuses to change tools.exec.ask or tools.exec.security; legacy tools.bash.* aliases normalize to the same protected exec paths.

Plugin-provided tools

Plugins can register additional tools. Some examples:

• Diffs — diff viewer and renderer
• LLM Task — JSON-only LLM step for structured output
• Lobster — typed workflow runtime with resumable approvals
• Music Generation — shared music_generate tool with workflow-backed providers
• OpenProse — markdown-first workflow orchestration
• Tokenjuice — compact noisy exec and bash tool results

Plugin tools are still authored with api.registerTool(...) and declared in the plugin manifest's contracts.tools list. OpenClaw captures the validated tool descriptor during discovery and caches it by plugin source and contract, so later tool planning can skip plugin runtime loading. Tool execution still loads the owning plugin and calls the live registered implementation.

Tool configuration

Allow and deny lists

Control which tools the agent can call via tools.allow / tools.deny in config. Deny always wins over allow.

{
  tools: {
    allow: ["group:fs", "browser", "web_search"],
    deny: ["exec"],
  },
}

OpenClaw fails closed when an explicit allowlist resolves to no callable tools. For example, tools.allow: ["query_db"] only works if a loaded plugin actually registers query_db. If no built-in, plugin, or bundled MCP tool matches the allowlist, the run stops before the model call instead of continuing as a text-only run that could hallucinate tool results.

Tool profiles

tools.profile sets a base allowlist before allow/deny is applied. Per-agent override: agents.list[].tools.profile.

coding includes lightweight web tools (web_search, web_fetch, x_search) but not the full browser-control tool. Browser automation can drive real sessions and logged-in profiles, so add it explicitly with tools.alsoAllow: ["browser"] or a per-agent agents.list[].tools.alsoAllow: ["browser"].

The coding and messaging profiles also allow configured bundle MCP tools under the plugin key bundle-mcp. Add tools.deny: ["bundle-mcp"] when you want a profile to keep its normal built-ins but hide all configured MCP tools. The minimal profile does not include bundle MCP tools.

Example (broadest tool surface by default):

{
  tools: {
    profile: "full",
  },
}

Tool groups

Use group:* shorthands in allow/deny lists:

sessions_history returns a bounded, safety-filtered recall view. It strips thinking tags, <relevant-memories> scaffolding, plain-text tool-call XML payloads (including <tool_call>...</tool_call>, <function_call>...</function_call>, <tool_calls>...</tool_calls>, <function_calls>...</function_calls>, and truncated tool-call blocks), downgraded tool-call scaffolding, leaked ASCII/full-width model control tokens, and malformed MiniMax tool-call XML from assistant text, then applies redaction/truncation and possible oversized-row placeholders instead of acting as a raw transcript dump.

Provider-specific restrictions

Use tools.byProvider to restrict tools for specific providers without changing global defaults:

{
  tools: {
    profile: "coding",
    byProvider: {
      "google-antigravity": { profile: "minimal" },
    },
  },
}