Openrouter Deep Research MCP Server

Research questions move from planning into parallel agent work, with model selection based on local embedding comparisons. A synthesis stage brings the findings together through consensus and citations. The research tool can queue a job and return its identifier or stream results synchronously; batch_research dispatches several queries together. Users can inspect progress with job_status, request best-effort cancellation with cancel_job, and continue an investigation through context-aware follow-ups. The agent entry point routes requests to research or retrieval and can chain tool calls.

Reports, jobs, and the knowledge graph persist across sessions by default, with PGlite and pgvector providing knowledge storage. Hybrid search combines BM25 with vector matching, while retrieve accesses indexed content or database records and query runs read-only SQL with bound parameters. Saved reports can be retrieved by identifier, and history lists recent reports. For exploring alternate research paths, session tools provide undo, redo, named checkpoints, and forks. Graph tools support traversal, clustering, and importance rankings. These workflows are available through MCP clients such as Claude, Jan AI, and Continue.