Topic
A shared MCP hub that lists all 40 tools to every connected agent is a convenience until the billing agent starts calling create_user. The fix isn't stricter prompts — it's a smaller tool list per agent, derived from the spec.
Read articleWhen you hand MCP gateway access to support and PMs, a team-wide key becomes the incident waiting to happen. Scope the key to the human, the role, and the three ops they actually use.
Read articleThe OpenAPI spec lists prod as the first server. Your agent reads it and calls prod on every test run. The fix is not another line in the system prompt.
Read articleShared stdio MCP configs look convenient until someone on the team leaves. The right model is per-teammate credentials, not one shared URL everyone pastes into Cursor.
Read articleWe hand-maintained a tools.json that matched our OpenAPI doc for about three weeks. Then someone renamed a parameter and the agent started sending the old name to an endpoint that no longer accepted it. The fix was not better discipline. It was letting the spec be the source of truth.
Read articlestdio means every engineer, every CI run, and every teammate's laptop spawns a Node process. A hosted HTTPS MCP URL means one config block, one process, one place to debug.
Read articlePasting a production bearer token into ~/.cursor/mcp.json means every engineer's laptop holds a credential that can call your billing API directly. The gateway should inject it; the agent client should never see it.
Read articleWhen you convert an entire OpenAPI doc into MCP tools, the agent sees 87 endpoints including the one that deletes a tenant. Per-operation exposure is the blast-radius control you actually want.
Read articleConverting an OpenAPI doc into an MCP server is easy. Keeping that Node process alive, rotated, and reachable from every engineer's laptop is the part that quietly becomes a second job.
Read articleEvery integration adds another MCP server. The agent now has 40 tools, and half the time it calls the wrong one. Tool count is not capability.
Read articleMost MCP servers load everything eagerly, dump huge schemas into the system prompt, and make the agent pay for tools it will never call. The fix is boring: lazy, narrow, and spec-driven.
Read articleThe default move when an agent gets something wrong is to paste more files into the context. But more files means more noise, and noise is what made it wrong in the first place.
Read article