Vercel: The MCP Server Now Takes Feedback From the Agent

Vercel: The MCP Server Now Takes Feedback From the Agent

Vercel’s changelog entry of October 23 adds a submit_feedback tool to the Vercel MCP server: a way to send bug reports, friction reports and feature requests to the Vercel team directly from an agent or harness. You ask in plain language, and the assistant attaches structured context describing the task, the expected behavior and what went wrong. The tool requires sign-in, and it “sends feedback only when you ask your agent to share it.” It is a short entry, with no numbers in it, and the idea is bigger than its length.

There is no marketing claim to discount. What is worth noticing is the design. The feedback channel has moved to where the friction happens: the agent that just hit a confusing error is the one best placed to describe it, with the context still in hand. The consent rule is the right one. An agent that filed reports on its own would be telemetry wearing a feedback costume; this one acts only on the user’s instruction. The open question is what Vercel does with a queue of agent-written reports, and the changelog does not say.

The example Vercel chose is the most telling part. The sample report reads: “The deployment error didn’t name the missing environment variable.” Agents read error messages literally, and an error that does not say what is wrong is a dead end for them in a way it is not for a person who can go and look. Vercel has six API pages in the catalog: the core Vercel API that deployments run through, Vercel Webhooks, and the AI Gateway surface of the Vercel Chat Completions API, Vercel Chats API, Vercel Models API and Vercel Embeddings API. The agentic access profile records six operations, four of them acting.

The measurement makes the example land. Vercel holds a Kin Score of 57.5, strong, with contract quality and developer ergonomics both at 63.1 and contract governance at 13.6, the lowest facet. Agent Readiness is 45.7, agent-ready: the MCP server, auth clarity, rate-limit signal, delegated identity and protected resource metadata are lit. Error semantics is unlit. So are idempotency, OpenAPI examples and reversibility. The first piece of feedback Vercel imagined its agents sending is a complaint about error semantics, which is the dimension the catalog cannot yet find in Vercel’s contracts. The tool is a good one. The contracts are where the reports will point.

← Traefik Puts Runtime Governance in Front of OCI Agents
WorkOS: Zero Trust Means Checking Every MCP Tool Call →