wse-dev-mcp-client

Local stdio ↔ HTTP MCP proxy for the Iceberg Data wse-dev MCP deployed on Cloud Run.

This is the client side of a split MCP architecture. The server (with the real WSE-ops tools — Firestore queries, job debugging, test enqueue, etc.) runs on Cloud Run under IAM control; this tiny package runs on each developer's machine as a stdio MCP server and forwards every request to Cloud Run with an auto-refreshed Google identity token.

Claude Code (stdio) → wse-dev-mcp-client (local) → Cloud Run (HTTP + ID token)
                             │
                             └─ On each call: gcloud auth print-identity-token
                                              (cached 55 min; invisible refresh)

No tokens to paste. No wrapper scripts to rerun. Access is gated by Cloud Run IAM (roles/run.invoker) — if your Google identity has it, the proxy works; if not, Cloud Run returns 403.

Prerequisites

• Python 3.11+
• gcloud CLI on your PATH, logged in with an identity that has run.invoker on the target MCP service.
  • One-time: gcloud auth login
• Claude Code (or any MCP-capable client)

Install

With pipx (recommended — keeps the proxy in its own isolated env):

pipx install wse-dev-mcp-client

Or with uv:

uv tool install wse-dev-mcp-client

Plain pip works too:

pip install --user wse-dev-mcp-client

Register with Claude Code

claude mcp add wse-dev wse-dev-mcp-client \
  --scope user \
  -e WSE_DEV_MCP_URL=https://your-cloud-run-mcp-url

Replace https://your-cloud-run-mcp-url with the Cloud Run URL of your deployed MCP server. For Iceberg Data developers, ask a teammate or run:

gcloud run services describe wse-dev-mcp \
  --region=us-west1 --project=iceflow \
  --format='value(status.url)'

Verify with claude mcp list — you should see wse-dev: ... ✓ Connected.

How it works

On each MCP call:

1. Claude Code writes a JSON-RPC message to the proxy's stdin.
2. The proxy consults its in-memory token cache (TTL 55 min). If expired or empty, it subprocesses gcloud auth print-identity-token to mint a fresh Google identity token.
3. The message is forwarded over HTTPS to the Cloud Run URL with Authorization: Bearer <id-token>.
4. Cloud Run IAM validates the token; if the caller has roles/run.invoker on the service, the MCP server handles the request and streams the response back.
5. The proxy pumps the response over stdout to Claude Code.

That's the whole package. ~150 lines of Python.

Failure modes

Configuration

Compatibility

• Python: 3.11, 3.12, 3.13, 3.14
• OS: Linux, macOS; Windows untested (should work with WSL)
• Claude Code: any version supporting stdio MCP transport (most recent clients)

License

MIT. © 2026 Iceberg Data LLC.

Related

• Model Context Protocol
• Anthropic Claude Code
• The server side lives in the private agent-global-config repo at .agents/mcp-servers/wse-dev/.