mcp

automation_file MCP server

Three ways to launch the MCP stdio server, in order of preference:

automation_file_mcp                                # installed console script
python -m automation_file mcp                      # CLI subcommand
python examples/mcp/run_mcp.py                     # standalone launcher

The server offers the fourteen semantic tools (file_read, file_copy, pipeline_run, ...), which are bound to a permission policy, and the bridge that exposes every registered FA_* action. For an AI client, give the semantic tools a root and switch the bridge off:

automation_file_mcp --root /srv/reports --no-bridge                 # read-only
automation_file_mcp --root /srv/reports --allow-write --no-bridge   # may create files

All three launch styles accept the same flags:

Flag	Default	Description
--name	automation_file	serverInfo.name returned at handshake
--version	1.0.0	serverInfo.version returned at handshake
--allowed-actions	(all)	Comma-separated allow list for the bridge (e.g. FA_file_checksum,FA_fast_find)
--no-bridge	(bridge on)	Offer only the semantic tools
--root	(none)	A location the semantic tools may work in: a storage URI or a local directory; repeatable
--allow-write	off	Let the semantic tools create files
--allow-overwrite	off	Let them replace an existing file (needs --allow-write)
--allow-delete	off	Let them delete; a move deletes its source (needs --allow-write)
--max-read-bytes	262144	Most bytes file_read returns in one call
--max-write-bytes	1048576	Largest content file_write accepts
--max-results	200	Most entries a listing or a search returns
--max-search-bytes	8388608	Most bytes one content search reads
--pipeline-dir	(memory)	Where pipeline_create keeps definitions
--pipeline-actions	(by permission)	Comma-separated actions a pipeline run through MCP may call
--tools	(all fourteen)	Comma-separated semantic tools to offer, or none

Claude Desktop

Edit claude_desktop_config.json:

• Windows — %APPDATA%\Claude\claude_desktop_config.json
• macOS — ~/Library/Application Support/Claude/claude_desktop_config.json

claude_desktop_config.json in this directory is a ready-to-copy sample: the semantic tools on one directory, the bridge narrowed to an allow list, the full bridge, and the standalone launcher. Pick the one that matches your install.

Manual smoke test

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' | automation_file_mcp

A single line reply containing serverInfo means the server is healthy.

Security

Use --no-bridge for an AI client. The semantic tools stay inside --root and are read-only until --allow-write; the bridge is not bound by either. The default registry includes FA_run_shell, FA_encrypt_file, and other high-privilege actions that an MCP host may invoke without prompting, so when the bridge stays on, narrow it with --allowed-actions.

The full manual, with the permission model and an example workflow, is docs/source/Eng/usage/mcp.rst.