mcp-tools-server

MCP server on the official Python SDK (FastMCP): four tools, stdio and streamable HTTP transports from one codebase, read-only SQL, a filesystem sandbox, offline test suite.

What this demonstrates

• MCP servers with the official SDK: tools, resources, schemas, lifecycle, both transports.

• Tool design for LLMs: names, argument schemas and docstrings are the interface the model reasons over, so descriptions carry the DB schema, examples and constraints on purpose.

• Least-privilege security controls: SQL locked to read-only via a SQLite authorizer rather than regex filtering (plus multi-statement injection rejection), and a filesystem sandbox built on resolve() + is_relative_to, with size caps and binary detection.

• Structured JSON logging to stderr: every tool call, rejection and startup is one machine-parseable line, shapes and sizes only — never SQL text, rows or file contents.

• Layering: tool logic is pure typed functions in app/tools/ with zero MCP imports; app/server.py registers thin wrappers.

Related MCP server: MCP Server

Architecture

flowchart LR
    CD["Claude Desktop / Inspector"] -- stdio --> S["FastMCP<br/>portfolio-tools"]
    ANY["any MCP client"] -- "streamable HTTP :8082/mcp" --> S
    S --> SW["search_web"] --> IDX[("search_index.json")]
    S --> QD["query_database"] --> DB[("demo.db<br/>SQLite, read-only")]
    S --> FS["read_file / list_dir"] --> DOCS[("data/docs<br/>sandbox")]

Tools

All four tools declare TypedDict results, so responses ship as machine-readable structuredContent (with an output schema) alongside the usual JSON text. The files in data/docs/ are also published as MCP resources — concrete docs://<file> entries plus a docs://{name} template; resource reads go through read_file's sandbox and size cap.

Demo DB (seeded on first query, ~25 rows of job-market data): companies(id, name, industry, city), vacancies(id, company_id, title, grade, salary_rub, stack), applications(id, vacancy_id, applied_at, status).

Quickstart

make install
make run-stdio      # stdio (default), for clients that spawn the server
make run-http       # streamable HTTP on :8082/mcp
# or: docker compose up --build

Debugging: npx @modelcontextprotocol/inspector python -m app.server.

Claude Desktop

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\), restart, and the four tools appear in a new chat:

{
  "mcpServers": {
    "portfolio-tools": {
      "command": "python",
      "args": ["-m", "app.server", "--transport", "stdio"],
      "env": {
        "PYTHONPATH": "/absolute/path/to/mcp-tools-server",
        "DATA_DIR": "/absolute/path/to/mcp-tools-server/data"
      }
    }
  }
}

Claude Desktop sets no working directory for servers, so both paths must be absolute; if deps live in a venv, use its interpreter as command.

Configuration

Read from the process environment, no dotenv (see .env.example): DATA_DIR — sandbox root holding search_index.json, demo.db and docs/ (default ./data); MCP_HOST / MCP_PORT — HTTP transport bind (default 0.0.0.0:8082); MCP_API_KEYS — comma-separated bearer keys required by the HTTP transport (default demo-key; stdio is not gated — its client is whoever spawned the process).

Notes

• search_web is an offline stub: keyword ranking over 14 curated entries, so demos and tests are deterministic. A real search API drops in behind the same contract.

• query_database layers, least-privilege: read-only URI (mode=ro), a sqlite3 authorizer that allowlists SELECT/READ/FUNCTION and denies writes, DDL, PRAGMA, ATTACH and transactions, complete_statement rejection of SELECT 1; DROP ... payloads, and row/cell/result-size caps.

• read_file / list_dir are sandboxed to DATA_DIR: absolute paths are accepted only when they resolve inside the root; any escape (.., outside path, or symlink target) is denied.

• Structured JSON logging (app/core/logging.py) writes one JSON object per line to stderr (stdout carries the stdio JSON-RPC frames), logging only call shapes/sizes/paths — never data.

• All failures raise ToolError with a single-line message — clients see actionable errors, never tracebacks.

Testing

106 tests, offline: units hit the pure functions directly (including the HTTP bearer gate through httpx.ASGITransport); integration runs a real MCP client against the server in memory via the SDK's create_connected_server_and_client_session — tools, structured output, resources and error mapping included. make install-dev && make test; make lint for ruff, make typecheck for strict mypy. CI runs lint + strict mypy + tests, a security job (pip-audit + bandit), and CodeQL; Dependabot keeps deps current. The seeded data/ (demo.db, docs/, search_index.json) is resolved from the repo root by default; set DATA_DIR to relocate the sandbox.

MIT. Portfolio demo — siblings: llm-gateway · rag-pgvector · agent-orchestrator, which calls this server's search_web over HTTP. Run the whole platform (all four services + web console + Postgres/Redis behind one origin) from the umbrella compose in llm-platform-console — docker compose up --build there brings everything up offline on mock models.

Releases

Version history is in CHANGELOG.md.