OpenRouter Agents MCP Server
@wheattoast11
A multi-agent research MCP server + mini client adapter - orchestrates a net of async agents or streaming swarm to conduct ensemble consensus-backed research. Each task builds its own indexed pglite database on the fly in web assembly. Includes semantic + hybrid search, SQL execu
概要
What is OpenRouter Agents MCP Server?
OpenRouter Agents MCP Server is a production-grade MCP server for multi-agent AI research and synthesis. It orchestrates parallel research agents using OpenRouter’s model API, providing planning, parallelization, and synthesis capabilities.
How to use OpenRouter Agents MCP Server?
Install via npx @terminals-tech/openrouter-agents --stdio (default STDIO transport) and set the OPENROUTER_API_KEY environment variable. Configuration can be provided via .env or .mcp.json file. For HTTP transport, add the --http flag.
Key features of OpenRouter Agents MCP Server
- Plans, parallelizes, and synthesizes multi-agent AI research.
- Embeds query vectors to route models without an LLM call.
- Persistent knowledge base with PGlite and pgvector.
- Circuit breaker for model API fault tolerance.
- Rail Protocol for bidirectional agent communication.
- Session time-travel with undo, redo, and fork.
Use cases of OpenRouter Agents MCP Server
- Decompose complex research queries into parallel sub-tasks.
- Synthesize results from multiple agents with citations.
- Persist and search research reports and knowledge graphs.
- Backend for AI assistants requiring multi-step reasoning.
- Integrate with MCP-compatible clients like Claude, Jan AI, and Continue.
FAQ from OpenRouter Agents MCP Server
What is the default transport?
STDIO is the default transport per MCP spec. Use the --http flag explicitly for HTTP mode.
Does it require an API key?
Yes, OPENROUTER_API_KEY is required. Key rotation and cooldown are supported via optional configuration variables.
Does data persist across sessions?
Yes, reports, jobs, and the knowledge graph persist by default. Set DB_AUTO_HEAL=true for in-memory mode (no persistence).
What models are supported?
Two tiers: high‑cost (e.g., Claude Sonnet 4.5, GPT‑5.2) and low‑cost (e.g., Gemini Flash, DeepSeek). Models are selected via embedding‑based routing.
Is it compliant with the MCP specification?
Yes, it complies with MCP Specification 2025‑11‑25 (Stable) and supports core features like Tools, Resources, Prompts, and SEP extensions.
