Eyes-MCP

Research MCP for AI agents. Docker-packaged. Swarm-ready. Zero required secrets.

A Model Context Protocol server that does the boring part of "go look this up on the internet" for any MCP-compatible agent. Callers send a question + a few parameters, Eyes-MCP fans out across SearXNG, Crawl4AI, GitHub, Reddit, YouTube, and more, and returns a synthesized answer with citations.

What we're building

flowchart TD
  Caller[Caller] -->|"prompt, depth, max_shards"| MainAgent
  subgraph MainAgent ["[main agent] - the only LLM"]
    MA1[1. Read available sources]
    MA2[2. Decompose prompt into N sub-tasks]
  end
  MainAgent -->|"N shards"| SubAgents
  subgraph SubAgents ["[N sub-agents] - async coroutines, no LLM"]
    SA1[search - SearXNG / GitHub / Reddit / YouTube / ...]
    SA2[fetch - Crawl4AI / API call]
    SA3[write raw output to disk - artifact pattern]
  end
  SubAgents -->|shard_id, path| Parse1
  subgraph Parse1 ["[parse layer 1] - bloat strip, depth-driven"]
    P1[strip HTML, nav, ads, boilerplate]
  end
  Parse1 --> Parse2
  subgraph Parse2 ["[parse layer 2] - chunk + per-shard dedup"]
    P2[chunk by structure then size cap 500 tokens]
    P3[simhash dedup within shard]
  end
  Parse2 -->|chunks grouped by shard_id| Review
  subgraph Review ["[main agent] - reviews and decides"]
    R1{synthesize or spawn more shards?}
  end
  Review -->|"synthesized answer"| Caller
  Review -->|"loop with gap-fill shards"| SubAgents

Decisions locked in:

• Only the main agent is an LLM. Sub-agents = async coroutines.
• Sub-agents write raw output to disk. Main agent sees pointers + chunks, not raw.
• Parse layer 1 = bloat strip. Aggressiveness driven by the depth param.
• No ranking. Every chunk goes to the main agent.
• Themes = shards. Main agent's decomposition defines the themes. No re-clustering.
• Crawl4AI's output is trusted for URL-fetched content — no re-stripping.

Why this is different

There are at least five "SearXNG + Crawl4AI + MCP" wrappers already (see 02-existing-landscape.md). What makes Eyes-MCP not the 6th me-too entry:

| Differentiator | What it means | |---|---| | Zero-friction install | docker compose up works with no API keys. LLM is optional. | | Shared service | Streamable HTTP, stateful sessions, multi-tenant. Designed to be a service, not a personal daemon. | | Source-aware adapters | GitHub, Reddit, YouTube, academic — each has a dedicated adapter, not a generic search-then-scrape path. | | Swarm-ready | Main agent + sub-agents + two parse layers wired from day one (swarm flag-gated in v1). | | Artifact pattern | Sub-agents write to disk, main agent reads chunks. Token cost is bounded by EYES_TOKEN_BUDGET. |

Install

git clone https://github.com/your-org/eyes-mcp.git
cd eyes-mcp
cp .env.example .env       # edit GEMINI_API_KEY if you have one
docker compose up

The first run pulls three images: searxng/searxng, unclecode/crawl4ai, redis. Build pulls node:20-alpine. Total cold start: ~2 minutes.

Important: before exposing SearXNG to the network, replace secret_key in searxng/settings.yml:

openssl rand -hex 32

Configure

All configuration is via environment variables. See .env.example for the full list with defaults. Highlights:

| Variable | Default | Purpose | |---|---|---| | EYES_HTTP_PORT | 8787 | HTTP listen port | | GEMINI_API_KEY | (empty) | Optional. If unset, server runs in heuristic-only mode. | | GEMINI_MODEL | gemma-4-31b-it | Model served by the Gemini-compatible API | | SEARXNG_URL | http://searxng:8080 | Internal SearXNG endpoint | | CRAWL4AI_URL | http://crawl4ai:11235 | Internal Crawl4AI endpoint | | GITHUB_TOKEN | (empty) | Bumps GitHub REST rate limit 60/hr → 5000/hr | | REDDIT_CLIENT_ID / REDDIT_CLIENT_SECRET | (empty) | PRAW auth (optional) | | EYES_MAX_SHARDS | 5 | Max shards per request (cap 20) | | EYES_MAX_ITERATIONS | 2 | Max refinement iterations (cap 5) | | EYES_TOKEN_BUDGET | 80000 | Token budget for the main agent's context | | EYES_TIME_BUDGET_SEC | 120 | Wall-clock budget per request |

Usage

The server speaks the MCP streamable HTTP transport on /mcp and a plain health probe on /health.

curl http://localhost:8787/health
# {"status":"ok","version":"0.1.0","uptime":12,"dependencies":{...}}

MCP clients that work out of the box

• Claude Desktop — add to claude_desktop_config.json:

  {
    "mcpServers": {
      "eyes": {
        "type": "http",
        "url": "http://localhost:8787/mcp"
      }
    }
  }
  

• Cursor — same config under MCP servers.
• MCP Inspector — point at http://localhost:8787/mcp to call tools interactively.
• Any other MCP client supporting streamable HTTP.

Project layout

Eyes-MCP/
├── src/
│   ├── index.ts          ← HTTP server, /health, /mcp, signal handling (THIS SUBAGENT)
│   ├── health.ts         ← dependency probe (THIS SUBAGENT)
│   ├── util/logger.ts    ← winston logger (THIS SUBAGENT)
│   ├── tools/            ← MCP tool registration (subagent B)
│   ├── main-agent/       ← orchestrator + LLM client  (subagent B)
│   ├── llm/              ← Gemini/Gemma client         (subagent B)
│   ├── dispatcher/       ← shard fan-out + scheduling  (subagent B)
│   ├── adapters/         ← GitHub/Reddit/YouTube/etc.  (subagent C)
│   └── parse/            ← bloat strip + chunking      (subagent C)
├── searxng/              ← SearXNG config
├── data/                 ← runtime artifacts (gitignored)
├── docker-compose.yml
├── Dockerfile
├── package.json
├── tsconfig.json
├── .env.example
├── LICENSE               ← MIT
└── README.md             ← you are here

Status

Build in progress. This repo is the result of a parallel subagent build:

• Subagent A (this) — skeleton, Docker, entrypoint, health, logger. ✅
• Subagent B — main agent, LLM client, dispatcher, tool registration. ⏳
• Subagent C — source adapters, parse layers. ⏳

After wire-up, see 03-architecture-main-and-swarm.md for the full design.

License

MIT. See LICENSE.