Blog

Feedback API: your AI agent can now report bugs and feature requests

Thomas SchiavoneThomas Schiavone
Feedback API blog cover

The first real report filed through Courier's Feedback API didn't come from a customer. It came from the AI agent writing the Feedback API's own docs.

The agent needed an example of a good report, so it went looking for real behavior to describe. It found some: GET /notifications/{id}/metrics returns 200 with an empty series when the template ID doesn't exist, while every other template endpoint returns 404. A typo in a template ID looks exactly like a template that stopped sending. The agent filed it with submit_feedback, and our team had it in real time.

Agents love to tell you what they find: the bug they hit, the error that could be clearer, the feature that would have saved a step. Until now, that feedback stopped with you. The Feedback API gives your agent a place to put it, so it reports what it finds and moves on. You can file your own through MCP or the API too.

AI agents find the rough edges first

A lot of the Courier code in your project wasn't typed by you. Your coding agent wrote the send, created the template, and built the journey. It reads the docs, calls the API, and reads the error when something's off.

That makes it the first to hit the rough spots, and it's better at finding them than most people. An agent reads every field in a response and checks it against the docs. It tries the edge cases a person would skip, and it doesn't give up on a confusing error after the third try. The docs agent found the metrics gap by comparing one endpoint against every other template endpoint, a check that rarely happens by hand.

What happens next is the problem. A person who hits a rough spot files a ticket. An agent tells you about it, then retries, works around it, and moves on. The workaround ships, and what it learned never reaches us.

We built Courier so people and agents work on the same platform. Everything in the dashboard is on the API, so an agent can build anything a person can. Hearing from those agents is the other half of that. If your agent burns ten attempts on one payload shape, we want to know.

What the Feedback API does

Courier's Feedback API lets AI agents, and the people working with them, report bugs, confusing behavior, and feature requests straight to Courier's team. It's one endpoint, POST /feedback, with four ways in:

Every report has a category (possible_bug, confusing_behavior, feature_request, or other), the surface where it happened (api, sdk, mcp, ui, docs, or other), a title, and a description. Optional fields carry the endpoint or page involved, what should have happened, the request ID of the call that shows the problem, and the model and agent that sent it.

Courier Skills tells your agent what to report, what to put in the report, and to tell you once it's sent. Here's the report the docs agent filed, lightly trimmed:

JSON
{
  "category": "confusing_behavior",
  "surface": "api",
  "title": "Template metrics return 200 with an empty series for an unknown template ID",
  "description": "GET /notifications/{id}/metrics returns 200 with an empty series when the template ID doesn't exist. The template metrics docs call this out as indistinguishable from a real template that has sent nothing. Every other template endpoint in the public spec returns 404 for an unknown ID: GET, PUT, and DELETE /notifications/{id}, and GET and PUT /notifications/{id}/content. In practice a typo in a template ID makes a dashboard show a template that stopped sending.",
  "location": "GET /notifications/{id}/metrics",
  "expected_behavior": "A 404 for an unknown template ID, matching the other template endpoints.",
  "model": "claude-opus-5-5",
  "harness": "Claude Code"
}

It states what happened, shows why it matters, and names the change it wants.

How we designed the Feedback API

Report anything that slows your agent down

Bugs count, and so does everything smaller: an error message that took three tries to decode, a docs page that sent your agent down the wrong path, a parameter you expected to exist. If it cost your agent time, send it. Repeats are welcome too. Hearing the same thing from ten agents tells us what to fix first.

Reports turn into pull requests

Courier's team and their agents read every piece of feedback in real time. Our agents review each report, check it against our code, and stage a pull request for our team to review, approve, and build on. A clear report from your agent can become a fix without anyone writing a ticket.

Your agent keeps working

Your agent doesn't stop to ask permission. It files the report, tells you in one line with the title and ID, and goes back to the task. You stay in the loop without losing your flow.

Model and harness point us to the right fix

Each report can name the model and the agent that sent it, such as claude-opus-5-5 and Claude Code. When one agent keeps tripping on the same parameter, that tells us whether the fix belongs in our API, our docs, or Courier Skills. Agents leave these out rather than guess.

Safe to leave on

You can turn it on and stop thinking about it:

  • Only an API key with no scope restrictions can send a report, so your end users can't report in your name.
  • Your agent sends each issue once per session, and resends a rejected report once at most.
  • It never sends messages or changes data to reproduce a problem.
  • Reports describe problems with IDs, never API keys, customer data, or your source code.

Turn on agent feedback

To let your agent report on its own, install Courier Skills:

Terminal
npx skills add trycourier/courier-skills

Then connect the MCP server. In Claude Code:

Terminal
claude mcp add --transport http courier https://mcp.courier.com --header api_key:YOUR_COURIER_API_KEY

The MCP server docs have setup for Cursor, Codex, and 12 more clients. Use an API key with no scope restrictions.

To file a report yourself, ask your agent through the MCP server, or run the CLI:

Terminal
courier feedback create \
  --category feature_request \
  --surface api \
  --title "Return rendered content in the send response" \
  --description "Goal: confirm template variables rendered after each send. POST /send returns only a requestId, so checking the output takes a second call to GET /messages/{message_id}/output." \
  --location "POST /send"

Wrapping up

Your agent already finds the rough edges in your Courier integration. Now what it finds reaches the people who can fix them, without you writing a ticket or your agent stopping its work. We're already looking at the first report.

Install Courier Skills, connect the MCP server, and let your agent tell us what it runs into. The Feedback API docs cover every field and limit.

Frequently asked questions

What is the Courier Feedback API?

The Courier Feedback API is an endpoint, POST /feedback, that lets AI agents, and the people working with them, report bugs, confusing behavior, and feature requests to Courier's team. Agents send reports through the Courier MCP server's submit_feedback tool, the Courier CLI, or the server SDKs, and Courier Skills tells them when and how to report.

What happens after my agent sends a feedback report?

Courier's team and their agents read every report in real time. Our agents can review it against our code and stage a pull request for our team to approve. For help with your own workspace, email support@courier.com.

Does my agent need the MCP server to send feedback?

The MCP server is one of four ways to send a report. Your agent can also use the Courier CLI, any of the seven server SDKs, or a POST /feedback request.

What shouldn't go in a feedback report?

API keys, secrets, customer data such as recipient emails or message content, and your own source code don't belong in a report. Describe the problem with IDs instead.

Which AI agents can report to Courier?

Any agent that can call an MCP tool, run the CLI, or make an HTTP request can send a report. Courier Skills installs in Claude Code, Cursor, Codex, and other coding agents that load agent skills.