Welcome to Jetson Orin NANO Developer Kit repository
06.OpenClaw Tools Introduction
6. Introduction to OpenClaw Tools
Tools are the "capability interfaces" that OpenClaw exposes to agents, allowing models to complete real operations in a safer and more structured way, such as reading and writing files, running commands, managing sessions, accessing web pages, and UI automation. Compared with the legacy openclaw-* Skills, tools are typed and controllable, and configuration can precisely restrict "what it is allowed to do / what it is not allowed to do".
6. Introduction to OpenClaw Tools1. Key Concepts1.1 The Relationship Between Tools and Skills1.2 More Tools Are Not Necessarily Better1.3 The Core of Tool Policy: profile + allow/deny2. Quick Start2.1 Selecting a Tool Configuration Profile2.2 Precise Control with allow/deny (Recommended: Start with deny)2.3 Tighten Only for a Specific Provider/Model (byProvider)3. Tool Groups (group:*)4. High-Risk Tools: How to Configure Them Safely4.1 Command Execution Related (group:runtime)4.2 Tips for apply_patch4.3 Isolate Tools by Agent (Recommended Practice)5. Recommended Configuration Templates (Ready to Use)5.1 Daily Dialogue (Safer)5.2 Studying/Writing Documents (Web Search Allowed, Running Commands Not Allowed)5.3 Development and Debugging (File Read/Write Allowed, but Running Commands Treated as "Optional")6. Frequently Asked Questions6.1 allow Was Configured but Does Not Take Effect6.2 One Model Can Use Tools, but Another Cannot6.3 Which Profile Should I Choose?7. Official Documentation
1. Key Concepts
1.1 The Relationship Between Tools and Skills
Tools: system-level "capability interfaces", typed, configurable, and restrictable by provider/model
Skills: more oriented toward "reusable workflows/plugin capabilities"; many legacy skills will be replaced or weakened by the tool system
If you want the agent to "do things", prioritize tools; if you want to encapsulate a process into a reusable capability, then consider Skills.
1.2 More Tools Are Not Necessarily Better
The more tools there are, the more the model can do, but the greater the risk. The best practice is: by default, enable only the tools you really need, and turn off high-risk capabilities (especially running commands) separately or enable them on demand.
1.3 The Core of Tool Policy: profile + allow/deny
The toolset is usually trimmed in this order:
tools.profilefirst provides a "basic allowed set"If
tools.byProvideris configured, the specified provider or provider/model is further tightened first (it can only narrow the toolset)Finally,
tools.allow/tools.denyare used to precisely tighten or loosen (withdenytaking precedence)
2. Quick Start
OpenClaw's tool allow policy is written in openclaw.json. In general, we do not need to configure it manually
denytakes precedence overallowThe
*wildcard is supported, and matching is case-insensitive
2.1 Selecting a Tool Configuration Profile
Common profiles (in order of "capability from least to most"):
minimal: onlysession_statusmessaging: message-oriented capabilities + sessions capabilitiescoding: development/engineering-oriented capabilities (files, runtime, sessions, memory, images, etc.)full: no restrictions (the same as leaving it unset)
Example: enable the coding profile
{ "tools": { "profile": "coding" }}2.2 Precise Control with allow/deny (Recommended: Start with deny)
Example: globally disable the browser tool
xxxxxxxxxx{ "tools": { "deny": ["browser"] }}Example: on top of coding, disable all runtime tools (exec / bash / process)
xxxxxxxxxx{ "tools": { "profile": "coding", "deny": ["group:runtime"] }}2.3 Tighten Only for a Specific Provider/Model (byProvider)
You can keep the global default unchanged and only further tighten the toolset for a specific provider or provider/model (it can only narrow, not expand).
Example: use coding globally, but minimal for Google Antigravity
xxxxxxxxxx{ "tools": { "profile": "coding", "byProvider": { "google-antigravity": { "profile": "minimal" } } }}Example: further reduce tools for openai/gpt-5.2
xxxxxxxxxx{ "tools": { "allow": ["group:fs", "group:runtime", "sessions_list"], "byProvider": { "openai/gpt-5.2": { "allow": ["group:fs", "sessions_list"] } } }}3. Tool Groups (group:*)
In tools.allow / tools.deny, you can use group:* to control a group of tools at once:
| Group Name | Expands to |
|---|---|
group:runtime | exec, bash, process |
group:fs | read, write, edit, apply_patch |
group:sessions | sessions_list, sessions_history, sessions_send, sessions_spawn, session_status |
group:memory | memory_search, memory_get |
group:web | web_search, web_fetch |
group:ui | browser, canvas |
group:automation | cron, gateway |
group:messaging | message |
group:nodes | nodes |
group:openclaw | All built-in OpenClaw tools (excluding provider plugins) |
Example: allow only file tools + browser
xxxxxxxxxx{ "tools": { "allow": ["group:fs", "browser"] }}4. High-Risk Tools: How to Configure Them Safely
4.1 Command Execution Related (group:runtime)
group:runtime contains execution capabilities (for example, exec / bash / process). For such capabilities, the following is recommended:
Disabled by default: first
deny: ["group:runtime"], then loosen it step by step when neededOpen only to "specific agents": concentrate high-risk capabilities on dedicated agents
Open only to "specific providers/models": grant more tools to stable/trusted models
4.2 Tips for apply_patch
apply_patch belongs to the file-editing tools; in some versions it is an experimental feature and may require an additional switch to enable (for example, via tools.exec.applyPatch.enabled, and only some models support it). If you find that a model clearly knows how to modify files but simply cannot call apply_patch, check the tool policy and the related switches first.
4.3 Isolate Tools by Agent (Recommended Practice)
You can set a global tool policy and then override it for a specific agent, so that "high-privilege capabilities" are concentrated on a small number of agents.
Example: use coding globally, but let the support agent use only messaging, and additionally allow Slack tools
xxxxxxxxxx{ "tools": { "profile": "coding" }, "agents": { "list": [ { "id": "support", "tools": { "profile": "messaging", "allow": ["slack"] } } ] }}5. Recommended Configuration Templates (Ready to Use)
5.1 Daily Dialogue (Safer)
Goal: chat only + view the session status only; do not give it the ability to "mess with the system".
xxxxxxxxxx{ "tools": { "profile": "minimal" }}5.2 Studying/Writing Documents (Web Search Allowed, Running Commands Not Allowed)
xxxxxxxxxx{ "tools": { "profile": "coding", "deny": ["group:runtime"], "allow": ["group:web"] }}5.3 Development and Debugging (File Read/Write Allowed, but Running Commands Treated as "Optional")
xxxxxxxxxx{ "tools": { "profile": "coding", "deny": ["group:runtime"] }}When you really need to execute commands, temporarily loosen group:runtime, and turn it back off after use.
6. Frequently Asked Questions
6.1 allow Was Configured but Does Not Take Effect
Check whether
denysimultaneously matches the same tool (deny takes precedence)Check whether the tool name is spelled correctly (matching is case-insensitive, but the name must exist)
If
tools.allowcontains only unknown/not-loaded plugin tools, the system may ignore the allow list to ensure that core tools remain available
6.2 One Model Can Use Tools, but Another Cannot
You may have set
tools.byProvider(provider or provider/model) to impose additional restrictions on that modelSome provider endpoints have inconsistent tool support; it is recommended to use a smaller toolset for unstable endpoints
6.3 Which Profile Should I Choose?
Just want to chat:
minimalDaily messaging + session management:
messagingFile/engineering tasks:
codingYou clearly know what you are doing and can bear the risk, and you want the complete, powerful OpenClaw:
full
7. Official Documentation
OpenClaw tool documentation: https://docs.openclaw.ai/zh-CN/tools