DB2ST MCP

Database-to-Structured-Tools MCP server. A horizontally scalable Model Context Protocol server that exposes shipment tracking (and, by design, other data sources) as authenticated MCP tools. Built in Python, deployable as a Knative Function.

The first tool wraps DB Schenker's public shipment tracking endpoint. The shape of the codebase is built so additional carriers, databases, or third-party endpoints slot in behind the same transport, auth, and scaling primitives.

Quickstart

uv sync --group dev              # install deps
cp .env.example .env             # configure
uv run db2st-mcp serve           # start HTTP server on :8080
uv run pytest                    # unit suite (integration tests deselected)
uv run pytest -m integration     # live upstream — needs unblocked egress IP
uv run pytest tests/e2e --report # E2E + Markdown report

Use as a local Claude Code MCP

# from the repo root
uv sync --group dev --extra fallback   # extra installs Playwright
uv run playwright install chromium     # download the browser driver
claude mcp add db2st-mcp -s user \
  -e TOKEN_STORE=memory \
  -e DB2ST_HTML_FALLBACK=1 \
  -- uv --directory "$(pwd)" run db2st-mcp stdio

# verify
claude mcp list | grep db2st-mcp   # should show: ✓ Connected

DB2ST_HTML_FALLBACK=1 enables the Playwright-based SPA scrape that takes over when the upstream JSON path is rate-limited — which it usually is from a fresh egress IP. Without the fallback, expect shipment not found on most first-call attempts. The [fallback] extra is what installs Playwright; see CONTRIBUTING.md for the full list of optional extras.

Then ask Claude Code to track a shipment, e.g. track DSV shipment 1806203236.

Two MCP tools are registered:

Tool	Returns
track_shipment	Full structured shipment (sender, receiver, package, history).
track_shipment_events	Just the chronological events timeline — lighter for poll-style clients.

Claude Code caches the MCP tool list per session. If you added this MCP before both tools were registered, restart Claude Code (or claude mcp remove db2st-mcp -s user and re-add) so the second tool is exposed. The same applies after pulling new code: the stdio subprocess holds whatever was on disk at claude mcp add time, so source changes (bug fixes, new tools) don't propagate until the next add.

Want to see the raw JSON-RPC handshake + a tools/call exchange without involving Claude Code? Run the example client:

uv run python scripts/example_call.py 1806203236

Use with other MCP clients

db2st-mcp is a vanilla stdio Model Context Protocol server, so any MCP-compatible client can run it. The package is on PyPI, which means the launch command across every client is the same:

uvx --from "db2st-mcp[fallback]" db2st-mcp stdio

This uses uv to download the package on first run, install the optional [fallback] extra (Playwright, needed when the upstream JSON API is rate-limited), and start the stdio transport. The first run will also need a one-time uvx playwright install chromium to download the browser driver.

If you don't want uv, swap uvx --from "db2st-mcp[fallback]" for pipx run --spec "db2st-mcp[fallback]" or any other tool runner.

The two environment variables the server respects are:

Var	Effect
TOKEN_STORE=memory	Skip the bearer-token check (suitable for local stdio; HTTP deployments should use the default upstash store + minted tokens).
DB2ST_HTML_FALLBACK=1	Engage the Playwright SPA fallback when the upstream JSON API 429s. Strongly recommended for local use.

Claude Desktop

Edit the config file (create it if missing):

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "db2st-mcp": {
      "command": "uvx",
      "args": ["--from", "db2st-mcp[fallback]", "db2st-mcp", "stdio"],
      "env": {
        "TOKEN_STORE": "memory",
        "DB2ST_HTML_FALLBACK": "1"
      }
    }
  }
}

Restart Claude Desktop to pick up the change.

Cursor

Per-project: .cursor/mcp.json in the repo root. Global: ~/.cursor/mcp.json.

Same JSON shape as Claude Desktop (the mcpServers object above).

VS Code (GitHub Copilot Agent)

VS Code uses a different schema — servers (not mcpServers) and a mandatory type field. Create .vscode/mcp.json:

{
  "servers": {
    "db2st-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "db2st-mcp[fallback]", "db2st-mcp", "stdio"],
      "env": {
        "TOKEN_STORE": "memory",
        "DB2ST_HTML_FALLBACK": "1"
      }
    }
  }
}

Continue (VS Code / JetBrains)

Create .continue/mcpServers/db2st-mcp.yaml in the workspace:

mcpServers:
  - name: db2st-mcp
    type: stdio
    command: uvx
    args:
      - "--from"
      - "db2st-mcp[fallback]"
      - "db2st-mcp"
      - "stdio"
    env:
      TOKEN_STORE: memory
      DB2ST_HTML_FALLBACK: "1"

MCP tools are exposed in Agent mode only.

Cline (VS Code extension)

Open the Cline panel → MCP Servers icon → Configure MCP Servers. That opens cline_mcp_settings.json (under VS Code's global extension storage). Add the same mcpServers block as Claude Desktop, plus Cline-specific keys if you want auto-approval:

{
  "mcpServers": {
    "db2st-mcp": {
      "command": "uvx",
      "args": ["--from", "db2st-mcp[fallback]", "db2st-mcp", "stdio"],
      "env": { "TOKEN_STORE": "memory", "DB2ST_HTML_FALLBACK": "1" },
      "disabled": false,
      "alwaysAllow": []
    }
  }
}

Zed

Zed keys MCP servers under context_servers (not mcpServers). Edit ~/.config/zed/settings.json:

{
  "context_servers": {
    "db2st-mcp": {
      "command": {
        "path": "uvx",
        "args": ["--from", "db2st-mcp[fallback]", "db2st-mcp", "stdio"]
      },
      "env": {
        "TOKEN_STORE": "memory",
        "DB2ST_HTML_FALLBACK": "1"
      }
    }
  }
}

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json (Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json). Same JSON shape as Claude Desktop.

Remote / HTTP-only clients

If your client doesn't speak stdio, point it at the deployed HTTP transport instead (see the next section for how to start one + mint a token). Most clients accept the same url + headers shape:

{
  "mcpServers": {
    "db2st-mcp": {
      "url": "https://your-host/mcp/",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Programmatic / custom clients

Any MCP SDK can drive db2st-mcp directly — there's nothing Claude-specific about the wire protocol. The tools/list → tools/call flow is plain JSON-RPC 2.0, and scripts/example_call.py is a 50-line reference implementation of a stdio client doing the handshake + calling track_shipment. Adapt that pattern for any agent framework (LangChain MCP adapter, LlamaIndex tools, your own runner) — the SDK choice is the consumer's; the server side stays unchanged.

Use as a deployed MCP (HTTP + auth)

# 1. start the server
uv run db2st-mcp serve

# 2. mint a token (one-time, secret is shown once)
uv run db2st-mcp mint --plan pro --limit 10000

# 3. call the MCP transport at /mcp with the bearer token.
#    Streamable HTTP requires the Accept header to list both types.

# 3a. Full shipment record:
curl https://your-host/mcp/ \
  -H "Authorization: Bearer $TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"track_shipment","arguments":{"reference":"1806203236"}}}'

# 3b. Events timeline only (lighter — for poll-style clients):
curl https://your-host/mcp/ \
  -H "Authorization: Bearer $TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"track_shipment_events","arguments":{"reference":"1806203236"}}}'

Architecture

See ARCHITECTURE.md for system architecture and domain documentation.

Testing

See docs/TESTING.md for the unit/integration/e2e split, coverage gates, and CI workflows.

Contributing

See CONTRIBUTING.md and CLAUDE.md.

License

MIT