AI bug fixing with MCP
Your AI coding agent is good at fixing bugs and bad at being told about them. It gets “the pricing button is broken on mobile” — a sentence with no element, no breakpoint, no console output — and spends its context window guessing at which file you meant. Layerdock hands it the actual report instead.
Layerdock publishes an MCP server for website feedback. Install it once next to your editor and Claude Code, Cursor or any Model Context Protocol client can read your bug reports directly: the element that was clicked, the CSS selector, the screenshot, and the console errors and failed network requests captured at that exact moment.
What your AI coding agent receives for one bug
The interesting part is not that the agent can list tickets — plenty of trackers expose an API. It is what one report contains by the time the agent reads it. Every Layerdock pin is rendered as a fix brief:
- Where it is — the page URL, the section, the CSS selector, and the component name when your framework exposed one.
- A search string that actually matches source — one contiguous run of literal text, chosen so a grep finds it. Not the element’s whole
innerText, which concatenates child nodes and embeds newlines that appear in no source file. - What the reporter asked for — their comment, in their words.
- The screenshot, passed as an actual image when one was captured. When none was, the brief says so plainly rather than sending the agent looking for it.
- The debug state at capture — console errors and failed or slow network requests, each stamped with how long before the report it happened. This is the $49/mo Agency plan; Pro sends element context and screenshots without it.
The five tools it exposes
| Tool | What it does |
|---|---|
list_docks | Lists the sites the agent can reach, with any codebase hints their owner declared — framework, source root, CMS. |
list_feedback | Answers “what needs my attention?” in one call — open work across every dock, or scoped to one site. |
get_feedback | The fix brief for one pin: location, a greppable search string, the reporter’s ask, the screenshot, and the console and network state at capture. |
update_feedback_status | Moves a pin through open → in_review → changes_requested → approved → resolved, within the role the token actually has. |
reply_to_feedback | Posts on the pin’s thread, attributed to the token owner and labelled as coming via an agent. |
The loop, end to end
A client pins feedback on your live site. They install the extension once and click the thing that is wrong — no account, no bug template, no guessing at what a developer needs. Layerdock captures the element, the screenshot and the diagnostics automatically.
In your editor you ask your agent what needs attention. It calls list_feedback, picks a pin, calls get_feedback for the full brief, greps your source with the search string it was given, and makes the change against the real console error rather than a paraphrase of it. Then it posts what it changed with reply_to_feedback and moves the pin with update_feedback_status.
The reviewer sees the reply on their own thread, labelled as coming via a coding agent, so a silent status flip never leaves them guessing. That labelling is stamped server-side from the authenticated token — an agent cannot claim to be a person.
Why no other feedback tool does this
Two things have to be true at once, and in this category they usually are not. The tool has to capture enough technical detail for an agent to act on — most stop at browser and OS metadata, or gate console capture behind their top tier. And it has to expose that over a protocol an agent can actually call, rather than a REST API you would have to wrap yourself.
Across the tools in our category survey, none ships an MCP server. That is the gap this fills, and it is the clearest reason to pick Layerdock over a more established competitor — BugHerd has a better task board and Marker.io has better tracker sync, and neither can hand a bug to your agent.
Which plan you need
- Pro, $19/mo — MCP access with element context and screenshots. Enough for the agent to find the code; not enough for it to diagnose a runtime failure.
- Agency, $49/mo — the same, plus the console and network bundle in every response. This is the tier that makes the agent useful on bugs nobody can reproduce.
Common questions
What is an MCP server for website feedback?
The Model Context Protocol is an open standard that lets an AI coding agent call external tools. An MCP server for website feedback exposes your bug reports as tools the agent can call directly, so it reads the actual report — element, selector, screenshot, console errors — instead of working from a description someone pasted into a chat window. Layerdock publishes one as @layerdock/mcp.
Which AI coding agents does it work with?
Any MCP client. Claude Code and Cursor are the ones most people use it with, and the setup guide covers both explicitly, but the server speaks standard MCP over stdio so anything implementing the protocol can connect — including your own client.
What does the agent actually receive for one bug?
A fix brief: where it is (page URL, section, CSS selector, and the component name when the framework exposed it), a search string chosen to actually match your source, what the reporter asked for, the screenshot as an image when one was captured, and the debug state at that moment — console errors and failed or slow network requests, each stamped with how long before capture it happened. The brief also states explicitly when no screenshot was captured, rather than sending the agent hunting for one that does not exist.
Can the agent close tickets on its own?
It can move a pin through open, in_review, changes_requested, approved and resolved, and post a reply on the thread — but only within the permissions of the token owner’s role on that dock; the server rejects anything else. Replies are attributed to the token owner and separately labelled as coming via a coding agent, so a reviewer can always tell an agent’s claim from a person’s.
Is the MCP integration on every plan?
Both paid plans have it, but they carry different payloads. Pro ($19/mo) sends element context and screenshots. Agency ($49/mo) additionally sends the console and network bundle, which is the part that lets an agent diagnose rather than guess. If handing full failure state to an agent is why you are here, that is the Agency plan.
Does it send my whole codebase anywhere?
No. The MCP server runs locally alongside your editor and reads feedback from Layerdock over an authenticated token. Your code never leaves your machine through it — the agent reads the bug report and works on your source locally, the same way it does with any other file.
Ready to wire it up? Connect your coding agent over MCP walks through the token, the editor config and the checks. See also what a bug report captures, what a dock is, or pricing. Layerdock is pre-launch — the pricing page takes waitlist signups rather than payment.