scouts-ai-mcp

Model Context Protocol (MCP) server that exposes the SCOUTS-AI web search API as a single web_search tool for AI agents, LLM apps, answer engines and GEO workflows.

• One tool, no API key. Backed by GET https://scouts-ai.com/api/search.

• Drop-in for Claude Desktop, Cursor, Open WebUI, Continue, Cline and any MCP host.

• Python ≥ 3.10, fastmcp (>=2.0, currently resolves to 3.x), httpx.

• MIT licensed.

Install

pip install scouts-ai-mcp

Related MCP server: OneSearch MCP Server

Run (stdio)

scouts-ai-mcp

That's it. Wire it into your MCP host of choice — for example, Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "scouts-ai": {
      "command": "scouts-ai-mcp"
    }
  }
}

Run (HTTP)

For remote MCP hosts and self-hosted bridges:

scouts-ai-mcp --transport http --host 127.0.0.1 --port 8765

Run (Docker)

A multi-arch (linux/amd64 + linux/arm64) image is published to Docker Hub as kecven/scouts-ai-mcp. It runs the MCP server in streamable HTTP mode on port 8765 and exposes the web_search tool at http://localhost:8765/mcp. No API key, no external dependencies.

Quick start:

docker run --rm -p 8765:8765 kecven/scouts-ai-mcp:0.1.7

Then point any streamable-HTTP MCP host at http://localhost:8765/mcp (or, when run behind a public proxy, https://<your-host>/mcp).

Override the upstream API base URL:

docker run --rm -p 8765:8765 \
  -e SCOUTS_AI_BASE_URL=https://scouts-ai.com \
  kecven/scouts-ai-mcp:0.1.7

Append CLI args (the image entrypoint is scouts-ai-mcp):

docker run --rm -p 8765:8765 kecven/scouts-ai-mcp:0.1.7 --log-level=DEBUG

Available tags:

• kecven/scouts-ai-mcp:0.1.7 — pinned, recommended for production.

• kecven/scouts-ai-mcp:0.1 — minor-version rolling tag.

• kecven/scouts-ai-mcp:latest — latest stable release.

Build and push locally (requires docker buildx):

docker buildx build \
  --platform linux/amd64,linux/arm64 \
  -f Dockerfile.hosted \
  -t kecven/scouts-ai-mcp:0.1.7 \
  -t kecven/scouts-ai-mcp:0.1 \
  -t kecven/scouts-ai-mcp:latest \
  --push .

The image runs as a non-root user (scouts, uid 1001) and includes a TCP-level HEALTHCHECK on 127.0.0.1:8765 (the streamable HTTP endpoint does not return 200 on plain GET, so an HTTP probe would be unreliable).

Tool: web_search

Returns a compact JSON object mirroring the SCOUTS-AI response shape:

{
  "query": "rust async runtime",
  "lang": "en",
  "page": 1,
  "pageSize": 10,
  "cached": false,
  "tookMs": 412,
  "results": [
    {
      "title": "Tokio - An asynchronous runtime for Rust",
      "url": "https://tokio.rs/",
      "content": "Tokio is an asynchronous runtime for the Rust programming language...",
      "publishedAt": "2025-11-14T00:00:00Z",
      "engine": "bing"
    }
  ]
}

Error handling

The tool raises ToolError (rendered as an MCP tool error) when:

• The query is empty/too long or lang/page are invalid → invalid arguments.

• The upstream returns 429 → rate limit exceeded; honors Retry-After when present.

• The upstream returns 5xx or the network call fails → SCOUTS-AI temporarily unavailable.

• The upstream returns a structured 4xx error envelope → forwards the code and message.

Configuration

All settings are environment variables. Defaults match the public SCOUTS-AI deployment.

Development

git clone https://github.com/kecven/scouts-ai-mcp.git
cd scouts-ai-mcp
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest

License

MIT — see LICENSE.