Skip to content

Hosted Deployment

The hosted path exposes the same read-only MCP server over Streamable HTTP. Local clients should continue to use stdio unless they are deliberately testing a container or hosted deployment.

Run this after Render deploys, Smithery listing changes, and release metadata updates:

  • Confirm Render uptime for the hosted service and inspect recent deploy logs.
  • GET / returns the status document and points clients to /mcp.
  • GET /healthz returns 200 with {"status": "ok"}; this is the /healthz platform health check.
  • GET /.well-known/mcp/server-card.json returns the current server name, version, homepage, icon, tool descriptions, read-only annotations, and output schemas.
  • /.well-known/mcp/server-card.json is the server-card metadata endpoint used for hosted discovery checks.
  • Confirm the Smithery listing points to the intended repository, branch, container build, and MCP endpoint.
  • Use manual MCP tool-call smoke in the Smithery Playground after release changes.

Manual MCP tool-call smoke should stay bounded:

list_economic_concepts(query="cash rate")
get_economic_series(concept="cash_rate_target", last_n=5)
get_derived_series(concept="real_cash_rate", last_n=5)
search_datasets(query="superannuation assets", source="apra")

The repository intentionally does not include scripts/mcp_http_smoke.py. Reintroduce automated hosted tool-call smoke only when the deployed transport and client naming are stable enough to avoid false release failures.

Vercel preview comments are docs-site signals. They confirm that the Astro/Starlight documentation site built and deployed for the relevant branch, including dependency-update branches that only touch lockfiles. They are useful for page rendering and documentation checks, but the signal is not MCP retrieval correctness evidence.

Treat Render health checks, Smithery metadata, server logs, catalogue audit output, integration tests, and manual MCP tool-call smoke as the evidence for hosted MCP correctness.

The public hosted server is acceptable because all tools are read-only and use public ABS, RBA, and APRA data. The main residual risk is large unauthenticated retrieval. Prefer examples with last_n, start, and end, specify table_id for APRA XLSX workbooks, and monitor request volume, CPU, memory, and upstream latency.