readonly-mcp-server

An MCP (Model Context Protocol) server with:

• Transport layer configuration — stdio for local process spawning, or HTTP+SSE for a networked/shared deployment. Selection follows "if stdio else http+sse": any MCP_TRANSPORT value other than the literal stdio falls back to HTTP+SSE.

• Connection configuration (authentication) — bearer API keys on HTTP+SSE (Authorization: Bearer <key>, required on the SSE connection and every message POST), and a pre-shared key on stdio so the same auth model applies to local connections too.

• Authorization limits (read / search only) — the server exposes exactly two tool categories, read and search. There are no write/update/delete tools at all, and each API key is independently granted a subset of {read, search}, enforced on every tool call.

Project layout

mcp-server-py/
├── config/
│   ├── api_keys.json     # key -> client_id + scopes (read/search)
│   └── documents.json    # sample read-only data backing the tools
├── src/
│   ├── config.py          # env-driven transport/auth/data settings
│   ├── auth.py            # API key store, TokenVerifier, scope checks
│   ├── data_store.py       # read-only document backend
│   ├── tools.py            # list_documents / read_document / search_documents
│   └── server.py           # wires transport + auth + tools, entry point
├── requirements.txt
├── .env.example
└── test_stdio_client.py / test_http_client.py   # example clients

Related MCP server: PDF MCP Server

Install

pip install -r requirements.txt
cp .env.example .env   # then edit as needed

Configuration (environment variables)

Edit config/api_keys.json to add/remove keys. Each key lists scopes, which may only contain read and/or search — anything else is rejected at startup.

Run — stdio transport

MCP_TRANSPORT=stdio MCP_STDIO_API_KEY=demo-read-search-key-change-me \
  python3 src/server.py

Point an MCP client (Claude Desktop, Claude Code, etc.) at this command, passing MCP_STDIO_API_KEY in the client's configured environment for the server process, e.g. in claude_desktop_config.json:

{
  "mcpServers": {
    "readonly-docs": {
      "command": "python3",
      "args": ["/absolute/path/to/mcp-server-py/src/server.py"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "MCP_STDIO_API_KEY": "demo-read-search-key-change-me"
      }
    }
  }
}

Run — HTTP+SSE transport

MCP_TRANSPORT=http MCP_HTTP_PORT=8000 python3 src/server.py

Clients connect to GET http://host:8000/sse and POST /messages/, both requiring Authorization: Bearer <api-key>. See test_http_client.py for a minimal client example using the official mcp Python SDK.

Tools exposed

A key missing the required scope gets a clear isError=true tool result (Access denied: this operation requires the 'X' scope...) rather than a crash or silent failure.

Design notes / how the auth is wired

• HTTP+SSE: uses the MCP Python SDK's built-in TokenVerifier + AuthSettings support. ApiKeyTokenVerifier (in auth.py) adapts our flat API-key file to that protocol; the SDK then wraps both the /sse and /messages endpoints with AuthenticationMiddleware + RequireAuthMiddleware automatically, and binds each SSE session to the credential that opened it (a POST with a different/no credential for an existing session is rejected).

• stdio: authenticated once at process startup via MCP_STDIO_API_KEY; scopes are held for the lifetime of the process. All logging goes to stderr — the stdio transport reserves stdout for JSON-RPC protocol frames.

• Scope enforcement: auth.require_scope() is called at the top of every tool function and reads the caller's granted scopes via get_access_token() (HTTP) or the stdio session's validated key (stdio), raising if the scope isn't present.