OpenRouter Deep Research MCP Server

The OpenRouter Deep Research MCP Server bridges the gap between conversational AI assistants and real‑world research workflows. It exposes a rich set of tools that let an assistant perform structured research tasks—planning, retrieving information, synthesizing results, and generating reports—while leveraging multiple LLMs on the OpenRouter platform. By packaging these capabilities into a single MCP endpoint, developers can embed sophisticated research behavior directly into their agents without managing separate orchestration logic.

At its core, the server offers two operating modes. AGENT mode presents a single tool that internally orchestrates the full research pipeline: it plans the steps, executes them in parallel (bounded by user‑defined limits), and synthesizes the final output. MANUAL mode splits the workflow into granular tools—, , , , and more—giving developers fine‑grained control over each phase. The default ALL mode combines both approaches and always includes lightweight operational utilities such as , , and job management. This duality lets teams start quickly with a single tool or evolve toward custom pipelines as their needs grow.

Key features that set this server apart include:

• Dynamic multi‑model orchestration: The server automatically selects from a catalog of cost‑effective and high‑performance models (Anthropic Sonnet‑4, OpenAI GPT‑5 family, Gemini, etc.) based on a declarative configuration. This enables cost‑aware planning and execution without hardcoding model choices.
• Bounded parallelism & workflow synthesis: The architecture allows the assistant to break complex tasks into sub‑steps, run them concurrently, and merge results. This boosts throughput while keeping resource usage predictable.
• Semantic knowledge base: Built on PGlite and pgvector, the server stores embeddings for fetched pages and user reports. It offers backup, export/import, health checks, and reindexing tools, making the knowledge base both durable and searchable.
• Lightweight web helpers: Quick search and page‑fetch tools provide contextual data on demand, feeding the LLMs with up‑to‑date information without external calls.
• Robust streaming & security: Server‑side events (SSE) support real‑time streaming of tool responses, and per‑connection authentication ensures only authorized clients can invoke the MCP endpoints.

In practice, this server powers research assistants that can autonomously conduct literature reviews, gather data from the web, summarize findings, and generate polished reports—all within a single conversational session. A product manager could use it to pull up the latest market analysis, while an academic researcher might leverage the semantic KB to compare papers quickly. By integrating seamlessly with MCP‑compatible clients (Cursor, VS Code, custom agents), developers can embed deep research capabilities into any AI workflow, from prototype prototypes to production‑grade services.