Laguarde MCP server

Laguarde

Laguarde is a self-hostable policy control plane for AI coding agents.

It gives a team one persistent place to define engineering practices, evaluate agent actions, ratify recurring developer preferences, and retain the exact policy revisions behind important decisions.

Laguarde runs locally for one developer or behind a team URL. Agents interact with the same server through standard MCP; humans use the dashboard and REST API.

What the prototype proves

initialization recipes, and PR review guidelines.

• Four policy categories: code rules, general guardrails, project

allowed.

• Four decisions: allowed, limited, approval, and forbidden.
• Context-specific policy bundles with immutable revision identifiers.
• One local daemon with a persistent registry of projects and Git origins.
• Fail-safe action evaluation: an unmatched action is limited, not silently

actions.

• Human approval for dependency, migration, deletion, and authentication

three observations promote it as a stronger candidate.

• SQLite decision records plus human-readable Markdown evidence.
• Developer feedback convergence: a human may merge any proposal immediately;

ratification.

• A dashboard for policy CRUD, decision evaluation/review, and feedback

Quick start

For a local MCP installation, use Node.js 24 or newer:

npx -y --package laguarde-mcp@0.3.2 laguarde-daemon ensure
npx -y --package laguarde-mcp@0.3.2 laguarde-daemon register --cwd .

The first command reuses the healthy local daemon or starts it once. The second registers the current Git repository and prints its project-specific MCP URL. All local projects share the daemon, dashboard, SQLite database, and audit history while remaining separately identifiable.

Conceptual MCP configuration:

{
  "mcpServers": {
    "laguarde": {
      "type": "http",
      "url": "http://127.0.0.1:3000/mcp/projects/RETURNED_PROJECT_ID"
    }
  }
}

For repository development, install Bun and run:

bun install
bun run build
bun run start

Project HTTP MCP endpoints live under http://localhost:3000/mcp/projects/:projectId, and agent-facing discovery is available at .

Onboarding surfaces:

• human guide: http://localhost:3000/guide;
• agent self-setup contract: http://localhost:3000/install (text/plain).

To onboard a capable agent, send it the /install URL and explicitly ask it to connect Laguarde for the current project. The contract tells it how to verify the server, make a minimal native MCP configuration change, discover the tools, and load the registered project's policy bundle.

Agent policy-gate skill

Install the optional fail-closed skill from this repository with:

npx skills add https://github.com/FuturPanda/laguarde --skill laguarde-policy-gate

The skill requires a cooperative agent to load the project-bound Laguarde policy bundle, evaluate and record every material action, and stop when policy is unavailable, limited, approval-required, or forbidden. It does not replace a sandbox or host-level execution hook.

For S3/CloudFront onboarding, generate the two static upload objects with:

bun run export:onboarding

See docs/s3-onboarding.md.

MCP Registry publication is automated through GitHub Actions after a one-time DNS authentication setup. See docs/registry-publishing.md.

The daemon's first start creates ~/.laguarde/laguarde.db, seeds global policy, and adds ten policies. Set LAGUARDEDATADIR, or the more specific LAGUARDEDBPATH and LAGUARDEEVIDENCEDIR, to place persistent data elsewhere.

Agent workflow

Shortened. The full README is on GitHub.