Vercel adds feedback command to CLI and MCP so agents can file bug reports

Vercel has given developers and their coding agents a direct line to its team. In two changelog entries published on the Vercel blog on the same day, the company introduced a vercel feedback command in its CLI and a matching submit_feedback tool in its Model Context Protocol (MCP) offering, both built to carry bug reports, friction reports and feature requests from the place where the work actually happens.

A feedback channel in the CLI

According to the Vercel blog, the new command takes a short message as an argument — a quick note that an error message failed to explain how to fix a problem, for instance. For richer reports, it supports structured context through a repeatable -f or --field flag that accepts key=value pairs. Vercel's own example attaches the product involved, the model and harness running an agent, the task being attempted, the expected outcome, the actual result, and a classification such as a bug.

The command also reads from standard input, so a longer write-up saved in a notes file can be piped directly into a report instead of being retyped. Getting the feature requires updating to the latest version of the Vercel CLI.

The same idea, built for agents

The MCP counterpart, submit_feedback, covers the growing share of Vercel usage that happens inside an AI assistant rather than at a human's prompt. Vercel describes a plain-language workflow: a user asks their agent to send feedback about, say, a deployment error that never named the missing environment variable, and the agent invokes the tool on the user's behalf.

Like the CLI command, the MCP tool can attach structured context — the tool that was being called, the task, the expected behavior, what actually happened, and a report type. Vercel notes two guardrails: the tool requires the user to be signed in, and it transmits feedback only when the user explicitly asks the agent to share it.

Why the structured fields matter

Both entry points share the same shape, and the key=value design looks deliberate. Free-form complaints are hard to triage at scale, but fields that separate the task from the expected and actual behavior map cleanly onto reproduction steps, and a type field lets incoming reports be sorted into bugs, friction and feature requests automatically. Because the fields are ordinary flags or JSON keys, an agent can populate them without a human writing a single line.

Why it matters

The distance between hitting a problem and reporting it is where most feedback dies. By placing a reporting command in the terminal — and, more notably, in the MCP surface that coding agents already use — Vercel reduces that cost to near zero for the people and programs most likely to run into rough edges.

The agent angle is the more consequential half. As more deployments and debugging sessions run through autonomous harnesses, tool failures occur in contexts no human ever observes. A feedback primitive the agent itself can call turns those sessions into a stream of actionable reports, gated by explicit user consent. It is also a template other platform vendors may follow: if MCP is becoming the standard interface between agents and services, feedback tools belong in that interface too.