Tools Reference

Cogtrix ships with 67 built-in tools across 8 categories. The agent loads tools on demand and you can pin or restrict them via the CLI or config.

Overview

Safety Categories

Category	Confirmation	Examples
Safe	No	`read_file`, `calculate`, `web_search`
Sensitive	Yes	`execute_shell_command`, `write_file`, `execute_python`

Confirmation Responses

When prompted for confirmation:

y — Yes, allow this execution once
n — No, deny execution
a — Allow all (approve tool for entire session)
d — Disable tool (block for this session)
f — Forbid all further tool requests (resets on next prompt)
c — Cancel the current agent workflow

On-Demand Loading

All tools start in an on-demand pool. The agent requests only the tools it needs for the current task through an internal request_tools meta-tool. When the agent calls request_tools, the requested tools are activated before the next turn. The agent can also release tools it no longer needs.

Agent-loaded tools are automatically unloaded at the start of each new prompt cycle so the LLM starts fresh. Tools loaded manually via /tools load, --activate-tools, or the API are pinned and persist across prompt cycles until explicitly unloaded with /tools unload.

You don’t need to manage this yourself — the agent decides which tools to load based on your request. See Tool loading for details.