Skip to main content

Agent tools

中文版

The model can use a bounded set of built-in tools. Approval mode, explicit allow-lists, and Hooks still decide whether a call runs.

Built-in tools

KindToolsTypical use
ReadOnlyread_file, read_many_files, grep, glob, list_directoryInspect files and paths
FileEditedit, write_file, save_memoryChange files or save a memory
ShellExecshellRun a shell command
Networkweb_fetchFetch an HTTP resource
Otherskill, todo, ask_user_questionReuse instructions, track work, ask the user

cosh_shell_evidence is available only when Core starts with --enable-shell-evidence-tool. Connected MCP tools use names such as mcp__<server>__<tool>; Extensions may add their own external names.

Approval choices

Core modeReadOnlyFileEditShell, network, MCP, external
trustRunRunRun
autoRunRunAsk
balanced, suggest, strictRunAskAsk

Unknown tool names are denied. The interactive shell maps recommend to strict approval, auto to auto, and trust to trust.

Limit what the model sees

Use --tools for exposure and --allowed-tools only for exact names that should bypass approval:

cosh-core --headless --tools read_file,grep,ask_user_question
cosh-core --headless --tools empty
cosh-core --headless --allowed-tools mcp__search__query

Allow-listing shell, a network tool, or an external tool grants real authority. Keep the list as narrow as the task requires.

Workspace and output limits

File-reading tools are rooted in the workspace captured when Core starts. A later shell cd does not change that boundary, and paths that escape it are rejected. Search results are bounded and report truncation when they are not complete. MCP output entering Agent context is limited to 64 KiB.

See MCP setup for external tools and Extensions for extension-provided tools.