OpenRouter Agents MCP Server

OpenRouter Agents MCP Server

[UPDATE – 2025-08-26] Two modes (set MODE env):

• AGENT: one simple tool (agent) that routes research / follow_up / retrieve / query
• MANUAL: individual tools for each action
• ALL (default): both AGENT and MANUAL, plus always-on ops tools

Diagram (simple)

• Killer features
  • Plan → parallelize → synthesize workflow with bounded parallelism
  • Dynamic model catalog; supports Anthropic Sonnet‑4 and OpenAI GPT‑5 family
  • Built‑in semantic KB (PGlite + pgvector) with backup, export/import, health, and reindex tools
  • Lightweight web helpers: quick search and page fetch for context
  • Robust streaming (SSE), per‑connection auth, clean logs

Install / Run

• Install (project dependency)

• Global install (optional)

• Run with npx (no install)

What’s new (v1.5.0)

• Version parity across npm, GitHub Releases, and GitHub Packages
• Dual publish workflow enabled

Changelog →

Quick start

1. Prereqs

• Node 18+ (20 LTS recommended), npm, Git, OpenRouter API key

2. Install

3. Configure (.env)

4. Run

• STDIO (for Cursor/VS Code MCP):

• HTTP/SSE (local daemon):

Windows PowerShell examples

• STDIO

• HTTP/SSE

One-liner demo scripts

Dev (HTTP/SSE):

STDIO (Cursor/VS Code):

MCP client JSON configuration (no manual start required)

You can register this server directly in MCP clients that support JSON server manifests.

Minimal examples:

1. STDIO transport (recommended for IDEs)

2. HTTP/SSE transport (daemon mode)

With the package installed globally (or via npx), MCP clients can spawn the server automatically. See your client’s docs for where to place this JSON (e.g., ~/.config/client/mcp.json).

Tools (high‑value)

• Always‑on (all modes): ping, get_server_status, job_status, get_job_status, cancel_job
• AGENT: agent (single entrypoint for research / follow_up / retrieve / query)
• MANUAL/ALL toolset: submit_research (async), conduct_research (sync/stream), research_follow_up, search (hybrid), retrieve (index/sql), query (SELECT), get_report_content, list_research_history
• Jobs: get_job_status, cancel_job
• Retrieval: search (hybrid BM25+vector with optional LLM rerank), retrieve (index/sql wrapper)
• SQL: query (SELECT‑only, optional explain)
• Knowledge base: get_past_research, list_research_history, get_report_content
• DB ops: backup_db (tar.gz), export_reports, import_reports, db_health, reindex_vectors
• Models: list_models
• Web: search_web, fetch_url
• Indexer: index_texts, index_url, search_index, index_status

Tool usage patterns (for LLMs)

Use tool_patterns resource to view JSON recipes describing effective chaining, e.g.:

• Search → Fetch → Research
• Async research: submit, stream via SSE /jobs/:id/events, then get report content

Notes

• Data lives locally under PGLITE_DATA_DIR (default ./researchAgentDB). Backups are tarballs in ./backups.
• Use list_models to discover current provider capabilities and ids.

Architecture at a glance

See docs/diagram-architecture.mmd (Mermaid). Render to SVG with Mermaid CLI if installed:

Or use the script:

If the image doesn’t render in your viewer, open docs/diagram-architecture-branded.svg directly.

Answer crystallization view

How it differs from typical “agent chains”:

• Not just hardcoded handoffs; the plan is computed, then parallel agents search, then a synthesis step reasons over consensus, contradictions, and gaps.
• The system indexes what it reads during research, so subsequent queries get faster/smarter.
• Guardrails shape attention: explicit URL citations, [Unverified] labelling, and confidence scoring.

Minimal‑token prompt strategy

• Compact mode strips preambles to essential constraints; everything else is inferred.
• Enforced rules: explicit URL citations, no guessing IDs/URLs, confidence labels.
• Short tool specs: use concise param names and rely on server defaults.

Common user journeys

• “Give me an executive briefing on MCP status as of July 2025.”
  • Server plans sub‑queries, fetches authoritative sources, synthesizes with citations.
  • Indexed outputs make related follow‑ups faster.
• “Find vision‑capable models and route images gracefully.”
  • /models discovered and filtered, router template generated, fallback to text models.
• “Compare orchestration patterns for bounded parallelism.”
  • Pulls OTel/Airflow/Temporal docs, produces a MECE synthesis and code pointers.

Cursor IDE usage

• Add this server in Cursor MCP settings pointing to node src/server/mcpServer.js --stdio.
• Use the new prompts (planning_prompt, synthesis_prompt) directly in Cursor to scaffold tasks.

FAQ (quick glance)

• How does it avoid hallucinations?
  • Strict citation rules, [Unverified] labels, retrieval of past work, on‑the‑fly indexing.
• Can I disable features?
  • Yes, via env flags listed above.
• Does it support streaming?
  • Yes, SSE for HTTP; stdio for MCP.

Command Map (quick reference)

• Start (stdio): npm run stdio
• Start (HTTP/SSE): npm start
• Run via npx (scoped): npx @terminals-tech/openrouter-agents --stdio
• Generate examples: npm run gen:examples
• List models: MCP list_models { refresh:false }
• Submit research (async): submit_research { q:"<query>", cost:"low", aud:"intermediate", fmt:"report", src:true }
• Track job: get_job_status { job_id:"..." }, cancel: cancel_job { job_id:"..." }
• Unified search: search { q:"<query>", k:10, scope:"both" }
• SQL (read‑only): query { sql:"SELECT ... WHERE id = $1", params:[1], explain:true }
• Get past research: get_past_research { query:"<query>", limit:5 }
• Index URL (if enabled): index_url { url:"https://..." }
• Micro UI (ghost): visit http://localhost:3002/ui to stream job events (SSE).

Package publishing

• Name: @terminals-tech/openrouter-agents
• Version: 1.3.2
• Bin: openrouter-agents
• Author: Tej Desai admin@terminals.tech
• Homepage: https://terminals.tech

Install and run without cloning:

Publish (scoped)

Validation – MSeeP (Multi‑Source Evidence & Evaluation Protocol)

• Citations enforced: explicit URLs, confidence tags; unknowns marked [Unverified].
• Cross‑model triangulation: plan fans out to multiple models; synthesis scores consensus vs contradictions.
• KB grounding: local hybrid index (BM25+vector) retrieves past work for cross‑checking.
• Human feedback: rate_research_report { rating, comment } stored to DB; drives follow‑ups.
• Reproducibility: export_reports + backup_db capture artifacts for audit.

Quality feedback loop

• Run examples: npm run gen:examples
• Review: list_research_history, get_report_content {reportId}
• Rate: rate_research_report { reportId, rating:1..5, comment }
• Improve retrieval: reindex_vectors, index_status, search_index { query }

Architecture diagram (branded)

• See docs/diagram-architecture-branded.svg (logo links to https://terminals.tech).

Stargazers