Skip to content

tools.md

Latest commit

 

History

History
130 lines (96 loc) · 5.98 KB

tools.md

File metadata and controls

130 lines (96 loc) · 5.98 KB
Raw

Tools

Tools are what turn a model into a coding agent: the agent decides what to call, Cortex runs it, and the result goes back into the conversation. Every call shows up in the timeline.

Tool names are case-sensitive and are what you pass to --enabled-tools and --disabled-tools, and what you list in an agent's tools field.

To see what a given run actually has available:

cortex exec --list-tools

Reading and searching

Tool Parameters Does
Read file_path, offset, limit Read a file. limit defaults to 2400 lines.
LS directory_path, ignorePatterns List a directory
Grep pattern, path, case_insensitive, line_numbers, context, context_before, context_after, glob_pattern, output_mode, head_limit Regular-expression content search. output_mode is file_paths or content.
Glob patterns, folder, exclude_patterns Find files by glob
SearchFiles pattern, path, content_pattern Fuzzy file search

Writing

Tool Parameters Does
Create file_path, content Create or overwrite a file
Edit file_path, old_str, new_str, change_all Replace text in a file
MultiEdit edits Apply several edits atomically
ApplyPatch patch, dry_run Apply a unified diff

Running commands

Tool Parameters Does
Execute command, workdir, timeout Run a shell command. command is an argument array.

Execute is the tool the sandbox and approval policies exist for. The default tool timeout is 900 seconds.

The web

Tool Parameters Does
WebSearch query, num_results, category, include_domains, exclude_domains, use_neural, livecrawl, type, context_max_characters Search the web
WebFetch url, format, timeout Fetch a page. format is text, markdown or html.
FetchUrl url, format, timeout Fetch a URL

Web search is enabled with --search.

Language intelligence

Tool Parameters Does
LspDiagnostics path, severity Diagnostics from the language server. severity is error, warning or all.
LspHover file, line, column Hover information. Positions are 1-based.
LspSymbols query, path Workspace symbol search

Planning and coordination

Tool Parameters Does
Plan title, description, tasks, agent_analyses, and optional architecture, tech_stack, use_cases, risks, success_criteria, timeline, estimated_changes Submit a structured plan for approval
UpdateGoal status (active, blocked, complete), optional progress, reason, evidence (kind = file/command/test + detail) Record evidence-based progress on the user's /goal. complete requires a reason and usable evidence. The model cannot pause.
ExitSpecMode reason Leave specification mode and unlock the mutating tools
Task mode (explore, plan, worker), prompt, description, context, await_result Delegate to a subagent
ListSubagents include_custom List the available subagent types
Questions title, questions Ask you a structured question. Question types are single, multiple, text and number.

See Plan and Spec modes and Long-horizon persisted goals.

Task tracking

Tool Parameters Does
TodoWrite todos Update the session todo list. Each item has id, content, status (pending, in_progress, completed) and priority (high, medium, low).
TodoRead — Read the current todo list

Skills and batching

Tool Parameters Does
UseSkill skill Load a skill into the context
Batch calls, timeout_secs, tool_timeout_secs Run 1–10 tool calls in parallel. Cannot nest Batch or call agent tools.

Tools from extensions

Source Naming
MCP servers mcp__<server>__<tool>
Plugins plugin_<slug>

What is available when

The tool set is not fixed. It narrows depending on context:

  • Child tasks cannot call Task, AskUser, Questions or send_to_user. Delegated work reports back through its parent instead.
  • Specification mode blocks the mutating tools until the agent calls ExitSpecMode.
  • Plan mode and read-only agents are restricted to the reading tools.
  • WebSearch and deep research are chat-surface tools.
  • --enabled-tools and --disabled-tools narrow the set further for a single exec run.
  • The permission table in config.toml can require approval for, or outright deny, individual capabilities.

Browser and desktop automation

Cortex ships no built-in browser or desktop-automation tool. Driving a browser is done by connecting an MCP server that provides those tools; the puppeteer entry in the MCP catalog is the one that does browser automation.

Because those tools arrive over MCP, they pass the same authority boundary as any other tool call: the sandbox, the approval prompt, and the permission deny list. /browser reports whether such a server is actually connected and names it — it never claims a capability the CLI does not have.

Computer is a different concept: it selects where tools run (Cloud, This PC, or SSH) via CORTEX_COMPUTER. It is not browser or desktop control.

See also