Use the MCP server

The Unyform MCP server lets your coding agent — Claude Code, Codex, or any MCP client — pull blueprints, conventions, and policy decisions straight from your org while it works. Where the gateway injects context automatically on every model request, the MCP server gives the agent tools it can call on demand: "search the blueprints," "validate this file," "is this action allowed?"

What it gives you

The server exposes 12 tools in four groups, each backed by the Unyform API:

• Blueprints — search_blueprints, get_blueprint, get_conventions
• Policy — check_policy, validate_output, get_policies
• Steering — request_approval, report_action, get_guidance
• Context — get_ecosystem_context, get_architecture, get_recent_decisions

Several context tools are stubs today and return a "not yet available" message — see the usage tour below for what's live.

Install

The server ships as the unyform-mcp binary:

unyform mcp status checks that the binary is on your PATH — your MCP client launches it by name.

Register it

The server speaks stdio by default (JSON-RPC over stdin/stdout), which is what every MCP client expects. Pass your org with --org-id.

claude mcp add unyform -- unyform-mcp --org-id <org-id>

{
  "mcpServers": {
    "unyform": {
      "command": "unyform-mcp",
      "args": ["--org-id", "<org-id>"]
    }
  }
}

[mcp_servers.unyform]
command = "unyform-mcp"
args = ["--org-id", "<org-id>"]

Configure

The server resolves three things from flags, environment, or project config — in that order:

Authenticate once with unyform login to write the credentials file, or set UNYFORM_API_KEY in the environment your client launches the server in. The key is sent as a bearer token; the server never embeds it in the config you paste into your editor.

Usage tour

Once registered, ask your agent to use the tools by name or by intent. A few that are live end to end:

• search_blueprints — "What blueprints cover error handling?" Returns the blueprints (id, name, description) whose name or description matches your query, scoped to your org.
• get_blueprint — "Show me the naming section of blueprint <id>." Fetches the full blueprint by id, or just one named section when you pass it.
• get_conventions — "What are our conventions for this file?" Returns your org's conventions compiled to markdown — naming, structure, testing, style.
• check_policy — "Can I add this dependency?" Submits an action and content and returns an ALLOW / BLOCK decision with the reason and the policy names that matched.
• validate_output — "Does this file pass?" Validates a file_path + content against your blueprints and policies, returning pass/fail with each violation's rule, message, severity, and suggestion.

The blueprints these tools read come from your repo analysis — see Codegraph for how that graph is built and what the blueprints capture.