Agent Tools

Talon’s AI agent comes equipped with a powerful set of built-in tools that let it take real action on your behalf — reading files, running commands, searching the web, and more. Each tool can be individually controlled through permission modes.

Available Tools

File Operations

The agent can read, write, and edit files directly in your workspace. This includes reading existing files for context, creating new files, and making targeted edits to specific sections without rewriting everything.

Read file: src/index.ts
Edit file: package.json (add dependency)
Write file: src/utils/helpers.ts (create new module)

Shell

Execute shell commands directly on your machine. Useful for running builds, tests, installs, or any CLI tool.

# Examples the agent can run on your behalf
npm install
git status
python manage.py migrate
docker compose up -d

Web Search

The agent can search the web and fetch content from URLs to pull in up-to-date information, docs, or research.

Search: "React 19 concurrent features"
Fetch: https://docs.example.com/api-reference

Memory

Store and retrieve persistent knowledge across conversations. The agent uses SQLite-backed memory to remember facts, context, and notes between sessions.

See Memory for full details.

Scheduling

Create scheduled tasks and reminders that the agent will act on automatically — even when you’re not actively in a conversation.

Remind me every Monday at 9am to review open PRs
Run database backup every night at midnight
Check API uptime every 5 minutes

Sub-agents

Spawn child agents to work on tasks in parallel. The parent agent coordinates results, letting complex workflows run concurrently instead of sequentially.

Task Management

Create, track, and complete structured tasks. The agent maintains a task list it can reference and update throughout a session.

Create task: Refactor authentication module
Update task #3: Mark as complete
List all open tasks

Config Management

Read and modify Talon’s runtime configuration without restarting. The agent can update settings, switch models, adjust permissions, and more — all on the fly.

Set default model to anthropic/claude-opus-4-5
Enable shell tool for this channel
Update system prompt

Permission Modes

Every tool is governed by a permission mode that controls whether the agent needs approval before acting.

Mode	Behavior
`plan`	Agent describes what it will do, waits for your go-ahead
`ask`	Agent asks before each tool use
`allow`	Agent proceeds automatically for trusted tools
`bypass`	No confirmation required (use with care)

You can configure permission modes globally or per-channel, and override them for specific tools. For example, you might set shell to ask while letting file reads run on allow.

# Example per-tool permission config
tools:
  shell: ask
  file_read: allow
  file_write: ask
  web_search: allow
  memory: allow