Vibe Coder MCP

Overview (confidence: medium)

Vibe Coder MCP is an open-source Model Context Protocol server (vibe-coder-mcp on npm, 100 stars, MIT) that extends MCP-compatible AI assistants (Cursor, Cline AI, Claude Desktop) with 10 specialized software-development tools. It provides research, planning, codebase analysis, PRD generation, user stories, task lists, fullstack scaffolding, and workflow execution.

Installed globally via npm install -g vibe-coder-mcp@latest (or run instantly with npx vibe-coder-mcp), it exposes a CLI (vibe) with interactive REPL mode and a setup wizard that configures the required OPENROUTER_API_KEY and project directories.

MCP Tools (confidence: high)

Ten tools are registered and auto-approvable by the client:

Tool	Purpose
vibe-task-manager	AI-native task management with Recursive Decomposition Design (RDD)
research-manager	Deep research via Perplexity Sonar (OpenRouter)
map-codebase	Advanced codebase analysis across 35+ languages
curate-context	Intelligent context curation for AI development
generate-prd	Product requirements document generator
generate-user-stories	User story generator
generate-task-list	Task list / development tasks generator
generate-fullstack-starter-kit	Fullstack project scaffolding
run-workflow	Multi-step workflow execution from workflows.json

Architecture (confidence: medium)

• Transport: stdio (default), SSE, WebSocket, HTTP — dynamic port allocation with conflict resolution.
• Semantic routing: embedding-based matching with sequential-thinking fallback; tool registry is centralized, self-registering.
• Session state: maintained across requests; unified agent coordination across all transports.
• Performance claims (from README): <200ms response, <400MB memory, 99.9% test success across 2,100+ tests; Code Map tool achieves 95–97% token reduction. Caveat: these are README figures; the repo marks illustrative samples as non-production; no external benchmark link captured.

Task Manager (confidence: medium)

vibe-task-manager implements Recursive Decomposition Design (RDD): 21 supported intents via multi-strategy recognition (pattern + LLM fallback), multi-transport agent coordination with capability mapping/load balancing, session persistence, real storage integrations (zero mocks), and artifact parsing from PRD/Task List generators.

Configuration (confidence: high)

Required: OPENROUTER_API_KEY (get from openrouter.ai). Unified project root via VIBE_PROJECT_ROOT (5-priority resolution chain, auto-detection for CLI) replaces legacy CODE_MAP_ALLOWED_DIR/VIBE_TASK_MANAGER_READ_DIR. .env template in src/config-templates/. MCP client config:

{
  "mcpServers": {
    "vibe-coder-mcp": {
      "command": "npx",
      "args": ["vibe-coder-mcp"],
      "env": {
        "OPENROUTER_API_KEY": "your-key",
        "VIBE_PROJECT_ROOT": "/path/to/project"
      }
    }
  }
}

autoApprove list covers all 9 generator tools plus map-codebase and run-workflow.

CLI & REPL (confidence: high)

• vibe — start MCP server (stdio)
• vibe --sse — start SSE server
• vibe "create PRD for todo app" — CLI natural-language mode
• vibe --interactive / vibe -i — REPL with session persistence, themes, markdown rendering, slash commands (/tools, /save, /export, /theme)
• vibe --setup — interactive wizard (configures OpenRouter key, directories, validates)
• vibe --reconfigure / --resume <id> / --list-sessions

Output organized under VibeCoderOutput/ (research/, prd-generator/, user-stories-generator/, task-list-generator/, fullstack-starter-kit-generator/, map-codebase/, vibe-task-manager/, workflow-runner/).

Related

openagentskill token-optimizer-mcp token-usage-reduction prompt-caching

Open Questions

• Whether the 99.9% / 95-97% figures hold under independent benchmark reproduction.
• How semantic routing via embeddings performs at scale vs the sequential-thinking fallback.

Sources

• raw/external/github-com-vibe-coder-mcp-afed2aee.md