openspec-mcp (Lumiaqian) — Summary

openspec-mcp is a TypeScript MCP server that exposes the full OpenSpec spec-driven development workflow as ~30 MCP tools plus an optional web dashboard, giving any MCP-capable AI client (Claude Code, Claude Desktop, Cursor) full programmatic access to OpenSpec's change, approval, review, and task management operations.

Problem it solves: The canonical OpenSpec CLI is powerful but requires the agent to issue shell commands; openspec-mcp wraps every OpenSpec operation as a first-class MCP tool so the AI can manage proposals, approvals, reviews, and task progress natively through tool calls rather than bash invocations, enabling richer multi-turn workflows.

Distinctive trait: Ships a complete web dashboard (--with-dashboard flag) on port 3000 (auto-incrementing) with a Kanban board, QA runner, approval queue, real-time WebSocket updates, and Markdown rendering — making it the only entry in this batch that pairs an MCP server with a fully featured web UI.

Target audience: Teams using the canonical OpenSpec CLI who want dashboard visibility, approval workflows, and programmatic AI access beyond what the CLI alone provides.

Production-readiness: v0.4.2, 24 stars, last commit January 2026. Published to npm as openspec-mcp. Requires @fission-ai/openspec CLI installed globally.

Relationship to seeds: Closest to ccmemory (Archetype 3 — MCP-anchored toolserver): like ccmemory, it wraps an external persistence layer (the OpenSpec file system) as MCP tools for AI consumption. Unlike ccmemory's Neo4j graph memory, openspec-mcp wraps file-based OpenSpec state. It also resembles taskmaster-ai (1 MCP server + file-based state), but adds a web dashboard and approval/review workflow tooling that taskmaster lacks.

openspec-mcp — Overview

Origin

Created by GitHub user Lumiaqian, a Chinese developer (the repo ships both English README.md and Chinese README.zh.md). Published to npm as openspec-mcp at version 0.4.2. The project is described as:

"MCP (Model Context Protocol) server for OpenSpec - spec-driven development with real-time dashboard and approval workflow."

Philosophy

The core thesis is that AI agents benefit from MCP-native access to OpenSpec operations rather than shell command invocations. The README highlights:

• MCP Tools: "Full OpenSpec CLI functionality exposed as MCP tools"
• Review System: First-class review comments with reply/resolve lifecycle
• Approval Workflow: Explicit draft → pending_approval → approved → implementing → completed → archived state machine
• Real-time Dashboard: Visual management interface with WebSocket live updates

Approval state machine (verbatim from README)

draft -> pending_approval -> approved -> implementing -> completed -> archived
                         -> rejected -> draft (revise and resubmit)

Available Prompts

The MCP server exposes two AI prompts (context-aware, for use in Claude/Codex):

Deployment modes

The server supports three deployment modes via CLI flags:

openspec-mcp — Architecture

Distribution

• Type: MCP server (npm package)
• npm name: openspec-mcp
• Version: 0.4.2
• Binary: openspec-mcp (defined in package.json#bin)
• Language: TypeScript (Node.js)
• Required runtime: Node.js >= 20.0.0

Dependencies (key)

Install commands

# Via Claude Code CLI (recommended)
claude mcp add openspec -- npx openspec-mcp

# With dashboard
claude mcp add openspec -- npx openspec-mcp --with-dashboard

# Claude Desktop / Cursor (mcp.json)
{
  "mcpServers": {
    "openspec": {
      "command": "npx",
      "args": ["-y", "openspec-mcp"]
    }
  }
}

Required external dependency

npm install -g @fission-ai/openspec

The MCP server wraps the OpenSpec CLI — openspec-mcp calls OpenSpec CLI commands internally via openspec-cli.ts.

Repository structure

openspec-mcp/
├── src/
│   ├── index.ts          ← MCP server entry point + CLI arg parsing
│   ├── core/
│   │   ├── approval-manager.ts
│   │   ├── context-analyzer.ts
│   │   ├── cross-service-manager.ts
│   │   ├── file-watcher.ts
│   │   ├── hooks-manager.ts
│   │   ├── openspec-cli.ts   ← Wraps @fission-ai/openspec CLI calls
│   │   ├── proposal-generator.ts
│   │   ├── review-manager.ts
│   │   ├── revision-manager.ts
│   │   ├── spec-critic.ts
│   │   ├── spec-parser.ts
│   │   ├── task-parser.ts
│   │   └── template-manager.ts
│   ├── api/              ← Fastify route handlers for dashboard
│   ├── server/           ← MCP tool registrations
│   ├── types/
│   └── utils/
└── web/                  ← Dashboard frontend (Vite/React/Tailwind)
    ├── src/
    ├── vite.config.ts
    └── tailwind.config.js

Target AI tools

• Claude Code CLI (stdio MCP)
• Claude Desktop
• Cursor
• Any MCP-compatible client

openspec-mcp — Components

MCP Servers (1)

Server name: openspec-mcp Protocol: stdio (MCP) Total tools: ~31

MCP Tools by Category

Guides & Context (4)

Management (4)

Validation (3)

Archive (1)

Tasks (4)

Approval (5)

Reviews (5)

Templates (3)

Generator (2)

Hooks (1)

Cross-Service (2)

MCP Prompts (2)

Web Dashboard Routes

Scripts

openspec-mcp — Uniqueness and Positioning

differs_from_seeds

The closest seeds are ccmemory (Archetype 3: MCP-anchored toolserver) and taskmaster-ai. Like ccmemory, openspec-mcp wraps an external state store as MCP tools for AI consumption — but where ccmemory uses Neo4j for knowledge graph memory, openspec-mcp wraps the OpenSpec file system. Like taskmaster-ai (1 MCP server + file-based tasks.json), openspec-mcp exposes task management via MCP — but adds approval/review workflow tooling, a web dashboard, and cross-service document management that taskmaster lacks. Unlike openspec (the reference seed, which ships commands + skills), openspec-mcp has zero Claude Code slash-commands or SKILL.md files — it is pure MCP. It is the only framework in this batch that combines an MCP toolserver with a local web dashboard in a single binary.

Positioning

• Niche: "OpenSpec + MCP + Web Dashboard in one binary"
• Primary differentiator: The only entry in this batch with a real-time Kanban web dashboard and approval queue management
• Secondary differentiator: Review comment system (add/reply/resolve) — treating AI-generated code changes with the same review primitives as human PRs

Observable failure modes

1. CLI dependency: Requires @fission-ai/openspec installed globally — if the CLI version is incompatible with what openspec-cli.ts expects, tools silently fail or return unexpected results
2. Single-project limitation: The server operates on one project directory at a time; multi-repo visibility requires running multiple instances
3. No authentication on dashboard: The web dashboard has no authentication mechanism — accessible to anyone on the local network if bound to a non-loopback interface
4. State drift: If the AI agent modifies openspec/ files directly (via Bash) while also using MCP tools, the file watcher may show stale state in the dashboard until the next chokidar event
5. No multi-agent coordination: Multiple agents calling the same MCP server simultaneously could produce race conditions on approval state files

Explicit antipatterns

None documented in README.