Server

@atomprd/mcp-server is the Model Context Protocol server for AtomPRD. It exposes 12 tools — 8 read, 4 write — over HTTP or stdio.

Why MCP?

A coding agent (Cursor, Claude Code, Cline, …) does not benefit from being handed a 40-page PRD. It benefits from being able to ask for the slice it needs:

• “Give me the criteria for feat_create_habit_h14a.”
• “What rules does this feature use?”
• “What’s the entity layout under module mod_habits_e5f6?”
• “Mark tsk_implement_create_h14a as done — I just shipped it.”

That’s what MCP is for. The agent makes a tool call; the server returns JSON.

Transports

# HTTP (default ATOMPRD_MCP_PORT=4100)
bunx @atomprd/mcp-server --project <project-id> --token <PAT> --transport http

# stdio (for IDE configs that prefer stdio)
bunx @atomprd/mcp-server --project <project-id> --token <PAT> --transport stdio

Authentication

The --token is a personal access token issued by the cloud. Tokens are scoped to a single project and have a default 90-day expiry. Rotate via the cloud dashboard.

The token grants access to the project’s atoms + relations. Write tools log to audit_log — every successful write is reviewable.

Configuration

The server reads project + token from CLI flags or environment:

Flag	Env	Default
--project	ATOMPRD_PROJECT	required
--token	ATOMPRD_TOKEN	required
--transport	ATOMPRD_TRANSPORT	stdio
--port	ATOMPRD_MCP_PORT	4100 (HTTP only)
--api-base	ATOMPRD_API_BASE	https://api.atomprd.dev

See also

• Tools — the 12-tool catalog.
• IDE integration — Cursor / Claude Code / Cline configs.
• Spec → Packages — package status.