Skip to main content

​
Tools

Tools are typed capabilities a selected Agent may call. They are not raw plugin config and they are not a second setup system. Normal setup starts from Agents, select the Agent, then open Agent > Tools. That page decides which discovered tools this Agent may use. Credentials and domain controls stay on the owning surfaces.

​
Setup Ownership

NeedUse
Allow or deny tools for an AgentAgent > Tools
Web/search, GitHub, Gmail, mediaAgent > Services
Chat delivery and channel actionsAgent > Channels
Skill instructions and depsAgent > Skills
Wallet approvals and signingWallets
Mining start/stop/capital/claimsMining
Marketplace offers and ordersMarketplace
Device pairing and diagnosticsAdvanced > Nodes
Agent > Tools should answer: “May this Agent use this capability?” It should not collect API keys, wallet approvals, mining capital, channel credentials, or plugin install settings.

​
Tool Families

FamilyExamplesMain docs
Runtime and filesexec, process, read, write, edit, apply_patchExec, Files
Web and browserweb_search, web_fetch, browserWeb, Browser
Messagingmessage, polls, reactions, threadsChannels
Sessions and Agentssessions_*, agents_list, subagents, ACPSubagents
Tasksscheduled task and run controlsAutomation
Nodes and mediacanvas, camera, screen, audio, location, notificationsNodes
Review and outputdiff_view, text-to-speech, reactions, thinking levelsDiff Viewer
Workflow helpersLLM Task and task/workflow orchestrationLLM Task, Tasks

​
Tool Policy

Tool policy can be global or per-Agent. 
{
  tools: {
    profile: "coding",
    deny: ["group:runtime"],
  },
  agents: {
    list: [
      {
        id: "support",
        tools: {
          profile: "messaging",
          allow: ["slack"],
        },
      },
    ],
  },
}
Profiles:
ProfileMeaning
minimalOnly basic session status.
codingFile, runtime, web, session, memory, and media tools.
messagingMessage and session tools.
fullNo profile restriction.
Groups:
GroupIncludes
group:runtimeexec, process, code_execution
group:fsread, write, edit, apply_patch
group:sessionssession list/history/send/spawn/yield/status and sub-agents
group:memorymemory search/read tools
group:webweb_search, web_fetch, x_search
group:uibrowser, canvas, diff view
group:automationtask/gateway automation controls
group:messagingmessage tool
group:nodespaired node tools
group:agentsagent list and plan update tools
group:mediaimage, music, video, and text-to-speech tools
group:fasedall built-in Fased tools except provider plugin tools
Provider-specific policy can only narrow access: 
{
  tools: {
    profile: "coding",
    byProvider: {
      anthropic: { profile: "minimal" },
    },
  },
}

​
Boundaries

  • A denied tool is not sent to the model provider.
  • A skill teaches the Agent how to use tools; it does not grant tool access.
  • A plugin may register tools; installing a plugin does not automatically grant every Agent access to those tools.
  • Wallet, mining, marketplace, and node actions keep their own approval and policy gates.
  • apply_patch is available only when enabled and allowed for compatible model paths.
  • Host or node exec needs approvals unless the operator deliberately changes that policy.

​
Start Here

Exec Tool

Shell execution, background processes, node execution, and approvals.

Web Tools

Search and fetch setup through Agent Services.

Browser

Managed browser profiles, snapshots, screenshots, and UI actions.

Skills

Instructions, dependency checks, templates, and per-Agent skill access.