Skip to main content

Tools reference

This page lists the tools implemented by coding_tools_mcp.server. Tool availability depends on --tool-profile.

Profiles

Server and workspace tools

set_default_cwd is not exposed by the read-only profile.

File and search tools

Notes:
  • read_file supports only utf-8 and rejects binary-looking files.
  • read_file.max_bytes is capped at 1 MiB.
  • search_text uses rg when available and falls back to Python scanning.
  • Default ignored names include .git, node_modules, build outputs, virtualenvs, and common cache directories.

Editing tool

apply_patch is hidden in read-only. Use dry_run to validate a patch without writing files.

Command tools

Command behavior is shaped by permission mode. In safe, network-looking commands, shell expansion, inline scripts, sensitive environment values, and broad temp writes are blocked. In trusted, local development workflows are allowed while secret filtering and Landlock remain active when available. In dangerous, command permission gates are disabled.

Git tools

These are Git inspection tools. They do not commit, checkout, reset, push, or mutate repository state.

Permission tool

The accepted tool_name values are exec_command and apply_patch. Permission names include network, destructive_command, long_timeout, sensitive_env, shell_expansion, inline_script, privileged_executable, and write_generated_or_ignored.

Image tool

view_image is enabled by default unless CODING_TOOLS_MCP_ENABLE_VIEW_IMAGE=0 or --enable-view-image is not used in a wrapper that disables it. The input file must be within the workspace path boundary.

Output convention

Tool calls return MCP content plus structured output. Successful tool payloads include ok: true where applicable. Failures are surfaced as structured errors with a code, message, category, retryability flag, and diagnostic details when available.