Tools¶

The agent exposes the following tools to the model. Each is offered only when it can run in the current environment (for example find and grep require ripgrep on PATH).

Tool	Purpose
`read`	Read the contents of a file, with optional line offset and limit
`write`	Write content to a file, creating it (and parent directories) as needed
`edit`	Edit a file by exact-text replacement, one or more targeted replacements at a time
`bash`	Execute a bash command in the current directory, with an optional timeout
`list_dir`	List the entries of a directory, one per line
`math`	Evaluate a Lua expression in a sandbox exposing the `math` and `bit32` libraries, and return its value
`webfetch`	Fetch a URL over HTTP(S) and return its content as markdown, text, or HTML
`kagi_search`	Search the web via the Kagi Search API and return the results as a numbered list (title, url, date, snippet). Offered only when a Kagi API key is set via `agent.configure_kagi`
`grep`	Search file contents for a pattern via ripgrep
`find`	Find files by glob pattern via ripgrep
`recall`	Retrieve the full original content of a compacted or earlier session item by its recall id
`search_history`	Search the recorded session (including content dropped by compaction) by regular expression or recall-id range, and return the matching recall ids
`plan_*`	Create and update a durable task ledger (a plan) and query what is ready to run. See Plans

Beyond the built-ins:

Defining your own tools covers writing a tool in Lua and offering it to the model.
Bridging an MCP server covers exposing an entire Model Context Protocol server's tools at once.
The tool_call hook covers intercepting a call before it runs, to rewrite its arguments or reject it outright.
Reducing token usage with rtk is a worked example of the tool_call hook that routes bash commands through rtk.