9 lessons from running a write-capable MCP server against 7 real APIs

Ecstatic-Hurry-635 · reddit · 2026-09-05

A developer running an MCP server with 15 tools that publish to LinkedIn, X, Instagram, Threads, Bluesky and YouTube Shorts shares production lessons for write-capable servers:

Ship a dry-run tool: a preflight tool checking char limits, unconnected profiles, missing media and credit cost became the most-called tool and boosted trust.
Partial failure is normal: store one post per profile so retries skip already-delivered targets instead of double-posting.
Validate the whole batch first: resolve all tool names in a 20-call multicall before executing any, so a typo in call 18 fails before call 1 writes.
Never default timezones silently: insist in tool descriptions and echo back the resolved absolute time so the model catches mistakes.
Expose health as a tool: getaccounthealth lets the agent debug failed posts without a web UI.
Auth from three sources: env vars (stdio), query params (remote connectors), headers (ChatGPT actions).
Don't make the model guess IDs: bind workspace to the API key; un-derivable required params become hallucinated values.
Tool descriptions are your prompt: rewording descriptions fixed more bugs than code changes.
OpenAPI shim gets ChatGPT almost free: keep handlers separate from the MCP transport layer and add a generated openapi. for Custom GPTs.

Original post →

9 lessons from running a write-capable MCP server against 7 real APIs

More from coding & agent