You shipped an MCP server a few weeks ago. It has a query tool, agents are calling it, and your dashboards are green. Every call returns a 200. Then a customer writes in to say the production numbers an agent gave them look odd, and you realize you have no way to see what the agent asked for or what it was handed. You don't control the agent, and you don't have access to the conversations.
That is the situation Braintrust starts from in its post, Understand how agents use your MCP server. The post puts it plainly: "You don't always control the agent using it, though, or have access to the conversations." A tool call can succeed and still give the agent the wrong data. Braintrust's example is a query tool that ignores an environment filter. An agent asks for production traffic, staging rows come back mixed in, and no error appears anywhere. That could be your odd customer report.
What Braintrust is offering is tracing for MCP calls. Each call is recorded with its inputs, results, timing and errors. It works through the JavaScript and Python SDKs, and it instruments both the official MCP SDKs and third-party implementations such as FastMCP. The APIs.io write-up notes that there are no marketing figures in the post, so the claim stands or falls on how it works for you.
Two optional additions are the part worth your attention. The first is a telemetry.intent field you can add to a tool's schema, which lets the agent say why it is making the call. The second is a send_feedback tool, which lets the agent report a confusing response along with the tool name and the arguments that produced it. Both turn the agent into a witness to its own session, so you hear about the confusion without having to be in the conversation. A feature called Patterns then clusters recurring problems across traces, so you are looking at groups of similar failures instead of one call at a time.
The advice that holds up even if you never touch Braintrust is simple. A 200 is not evidence that the agent got what it asked for. Record the request next to the answer, and you can check one against the other.
If you did want to follow this through Braintrust's APIs, the APIs.io catalog record covers 29 API pages, and a few of them fit this story. Traced calls are read in Logs, which is the Braintrust Logs API. The evals the post suggests building from representative calls run through the Braintrust Evals API and the Braintrust Datasets API. The servers themselves are registered through the Braintrust MCP Servers API. The agentic access profile counts 208 operations, 89 of them acting.
Here the catalog turns on the company itself. Braintrust's Kin Score is 48.0, rated developing, and the facet dragging it down is operational transparency at 23.7. That facet measures whether a provider publishes its own status, changelog and limits. It is the lowest facet on a company that sells visibility into other people's systems. Contract governance sits lower still, at 13.6.
There is better news too. Agent Readiness is 52.4, with an MCP server, delegated identity, protected resource metadata, dynamic client registration and OpenAPI examples all lit. But error semantics is unlit, and that is the exact gap the post is about. When a contract does not describe what wrong looks like, an agent cannot tell a wrong answer from a right one.
So you come away with two things. For your own server, trace the request alongside the response, and give agents a way to tell you when something looked off, because a green status code won't. And when you judge a vendor, including one that sells observability, check what it publishes about itself, since the catalog filed Braintrust's own transparency as its weakest facet.
Read the original: https://apis.io/2026/11/16/braintrust-traces-mcp-servers-and-scores-23-on-operational-transparency/