Claude Code / MCP

lgit mcp speaks the Model Context Protocol over stdio. Registering it gives an agent read and write access to the same findings store the CLI and the VS Code extension use, so an agent can run the gates, read what came back, and ask for a remediation recipe without you copying anything between windows.

Register the server

make install-mcp

That is a thin wrapper around:

claude mcp add -s user l0-git /absolute/path/to/lgit mcp

-s user matters: without it Claude Code registers the server for the current directory only, and the l0-git tools vanish in every other project.

Or write it into ~/.claude.json yourself:

{
  "mcpServers": {
    "l0-git": {
      "command": "/absolute/path/to/lgit",
      "args": ["mcp"]
    }
  }
}

Any MCP client works: there is nothing Claude-specific in the protocol surface. Check the registration with:

make status

Tools

Tool	Arguments	What it does
gates_check	project, gate_id?	Run all gates, or one, against a project root and persist the results
gates_list	-	The registered gate set: id, title, description, severity, tags
findings_list	project?, status?, severity?, gate?, tag?, query?, sort?, limit?, offset?	Filter, sort and paginate the findings store
findings_stats	project?	by_severity, by_status, by_gate, top_files, by_tag, a 7-day trend, and how current they are: last_checked_at and project_exists (with a project), projects_tracked and projects_missing (without), only a full check counts
findings_ignore	id	Mark a finding ignored so later runs do not resurface it
findings_delete	id	Drop a single finding
findings_clear	project	Wipe every finding for a project
findings_remediate	id	The structured remediation for one finding

What findings_remediate returns

This is the tool worth understanding, because it is the one that decides how much the agent should trust itself.

• summary: what needs to happen, in one line.
• confidence: deterministic or guided.
• recipe: for the eight deterministic gates: exact commands, exact file edits, and the caveats. Empty for everything else.
• claude_prompt: a self-contained prompt framing the fix, always present.

deterministic means the fix is mechanical and the recipe can be applied as written. guided means the gate found something real but the fix needs judgement, which image tag to pin, whether a credential is live, whether a placeholder is still wanted.

The server never executes anything

findings_remediate returns text. Applying it goes through the agent's own tools and its own permission prompts. l0-git deliberately has no write path into your working tree.

A typical loop

1. gates_check with the project root.
2. findings_list filtered to severity=error.
3. findings_remediate on each id.
4. The agent applies what it can and asks about the rest.
5. gates_check again to confirm the findings resolved.

Step 5 matters: findings are keyed by (project, gate_id, file_path), so a re-run flips a fixed finding to resolved rather than leaving a stale row behind.

Running the server from the extension

The VS Code extension can spawn lgit mcp for you: command l0-git: Start MCP server (manual), or the l0-git.autoStartMCP setting.

You usually do not want this. Claude Code spawns its own copy of the server from the registration above, and a second process against the same SQLite file buys you nothing. The setting exists for clients that expect to attach to an already-running server.