Tools

A tool is the unit the agent loop calls when the model decides it needs to act in the world. { name, description, parameters, execute }. The model picks a tool based on its instructions, the user's message, and the tools available to that agent.

TAI tools vs MCP. If you've used Model Context Protocol, the mental model is the same: structured JSON-Schema parameters, structured results. TAI doesn't speak MCP over JSON-RPC; tools are in-process TypeScript objects. The simplification is intentional: no separate server process, no transport, no version negotiation. The cost is that sharing tools across stacks is a copy-paste job until the plugin loader ships in v0.2.

How to enable

Each tool has a tools.<name> block in config.yaml. Set enabled: true to turn it on:

yaml

tools:
  exec:
    enabled: true
    allowedCommands: [git, pnpm, node, tsc]
  write:
    enabled: true
    allowedPrefixes: [data/, output/]
  web_search:
    enabled: true
    provider: brave
    apiKey: ${BRAVE_API_KEY}
  browser_mediator:
    enabled: true
    egressAllowList: [amazon.com, wikipedia.org]
    vaultEnabled: true
  discord_dm:
    enabled: true

To restrict which agents can call a tool, list it in the agent's tools: array. An agent with no tools: list inherits every enabled tool.

Built-in catalog

Memory and knowledge

Read Memory for how these stores relate.

Tool	Verbs	Notes
`memory`	`read`, `write`, `append`, `list`, `search`	File-shaped notes in the agent's context dir. Scopes: `profile`, `global`, `knowledge`.
`recall`	`note`, `query`, `search`, `list`, `delete`	The structured tier. Short notes with tags and embeddings.
`core_memory`	`append`, `replace`, `read`, `clear`	Always-injected text the agent maintains as its working summary.
`facts`	`set`, `get`, `delete`, `query`	Structured `(category, entity, key, value)` facts.

File system and shell

Tool	What it does	Notes
`read`	Read a file's contents.	`allowedPrefixes:` config to scope.
`write`	Create or overwrite a file.	`allowedPrefixes:` config.
`exec`	Run a shell command.	`allowedCommands:` allowlist. Honors sandbox config.

Web

Tool	What it does	Notes
`web_fetch`	Fetch a URL, extract plain text.	Aware of browser-mediator sessions: refuses when an active mediator session disallows the host.
`web_search`	Brave Search API.	Requires `apiKey: ${BRAVE_API_KEY}`.

Tasks and projects

Read Tasks & projects for the backend story.

Tool	What it does
`tasks`	Create, update, delete, comment on project tasks.
`task_query`	Filter tasks by status, assignee, tags, text.
`projects`	List, register, switch projects.
`documents`	Document store. PDF and image OCR via optional peers.

Integrations (Google)

Tool	Verbs	Setup
`gmail`	`search`, `read`, `mark_seen`, `send`	Requires the `gog` CLI installed and `${GOG_KEYRING_PASSWORD}` set. Account in `tools.gmail.account`. Send is opt-in per agent (add `gmail` to the agent's `tools:` allowlist deliberately).
`google_calendar`	`list_events`, `search`, `create_event`, `delete_event`	Same `gog` CLI.
`google_drive`	`upload`, `list`, `search`	Same `gog` CLI.

Browser

Tool	What it does	Notes
`browser`	Generic Playwright browser. Navigate, click, type, screenshot.	Requires `playwright` peer dep.
`browser_mediator`	Bounded browser with egress allow-list, vault refs, output sanitiser, always-HITL gates.	Provided by `@tailored-ai/browser-mediator`. Prefer this over `browser` for agent-driven web flows.

Delegation and control

Tool	What it does
`delegate`	Call another agent synchronously (or `async=true` for background).
`task_status`	List or inspect background tasks started by async `delegate`.
`claude_code`	Delegate to the Claude Code CLI as a sub-agent.
`ask_user`	Block the loop and prompt the user for input. Works in CLI and Discord.
`load_skill`	Lazy-load a Skill into the current loop.
`run_workflow`	Trigger a Workflow.

Outbound push

Tool	What it does
`discord_dm`	Send a Discord DM to the configured owner. The right channel for unsolicited "FYI" pings from online agents.
`request_action`	Enqueue a high-risk action (purchase, form submit) to the trusted-actions gateway for human approval.
`check_action_status`	Poll a previously-enqueued action's status.
`md_to_pdf`	Convert a markdown file to PDF. Requires `md-to-pdf` peer dep.

Admin and introspection

Tool	What it does
`admin`	Read or update agent configuration at runtime. Restricted to write-permitted agents.
`resource_admin`	Install, list, remove skills and workflows.
`sleep`	Block the loop for N seconds. Useful in workflow steps.
`current_datetime`	Returns the current local datetime.
`uptime`	Returns the agent process uptime.
`git_log`	Returns recent git log entries from the current project.
`ip_info`	Returns the agent's public IP and geo.

Restricting per agent

By default an agent inherits every enabled tool. Narrow with the agent's tools: array:

yaml

agents:
  researcher:
    tools: [web_search, web_fetch, recall, memory]
    # exec, write, browser, etc. are unavailable to this agent

For a 30B local model: keep the list under five tools per agent. Past that, accuracy of tool selection drops noticeably.

Custom tools

For tools that wrap a shell command, no TypeScript needed. The custom_tools: block in config.yaml takes a name, a command, a description, and a parameter schema. See Custom tools.

Plugin tools

Tools shipped as npm packages plug in via the plugins: config block (v0.2, see Plugins).

Writing tools in code

See Extending in code for the Tool interface and how to add a tool by forking the monorepo or by writing a plugin.

The Tool interface

For reference:

import type { Tool, ToolContext, ToolResult } from "@tailored-ai/core";

export const myTool: Tool = {
  name: "my_tool",
  description: "Does something useful in one sentence.",
  parameters: {
    type: "object",
    properties: {
      input: { type: "string", description: "The input." },
    },
    required: ["input"],
  },
  async execute(args, context: ToolContext): Promise<ToolResult> {
    return { success: true, output: `Result: ${args.input}` };
  },
};

ToolContext gives access to the active session id, project id, db reference, and runtime accessor. The optional destroy?() hook fires on hot reload, when the tool is being replaced.