Connect your coding agent to Cobalt

This page is part of the guides hub on connecting visual feedback to AI coding agents. It covers the highest-fidelity path: connecting your agent directly to Cobalt over MCP so it ingests the annotated screenshots as images. For the file-in-repo workflow, see getting feedback into Claude Code.

What you'll learn

• What the Cobalt MCP server is and why it exists
• How to add it to Claude Code, Cursor, or Codex
• What your agent receives when it reads a review
• The no-setup fallback for agents that only browse the web

The problem MCP solves

When you hand a coding agent a review as a plain link, most agents extract the text of the page and never open the images. But a Cobalt review is visual — a screenshot with an arrow pointing at the button that's misaligned, a label that says "this is cut off." If the agent can't see the picture, it's working from half the feedback.

The Model Context Protocol (MCP) fixes this. Instead of hoping the agent fetches the image, the MCP server hands it over directly, as a native image block the agent's vision reads the same way it would read a screenshot you dragged into the chat.

Add the server (once)

For Claude Code:

claude mcp add --transport http cobalt https://cobaltcapture.com/mcp

Cursor, Codex, and other MCP clients take the same endpoint — https://cobaltcapture.com/mcp — over the Streamable HTTP transport. There's no account, no API key, and nothing to configure: the server serves public reviews with the same access as opening the link in a browser.

Then ask your agent

Once connected, point the agent at any review:

Read the Cobalt review https://cobaltcapture.com/r/abc12345 and make
the changes it asks for.

The agent calls the get_review tool with that URL. Back comes the review's markdown — the title, your notes, and a "Marked on the image" line describing each annotation — followed by every screenshot as an image/png block with the arrows, boxes, pins, and labels burned in. The agent sees exactly what your reviewer marked and works through the findings.

The no-setup fallback

Don't want to add a server, or using an agent that doesn't speak MCP? Just paste the link. When an AI fetcher opens a Cobalt review URL, the server returns the markdown export instead of the HTML page — with the annotations described in text and an explicit note telling the agent to fetch and view each screenshot. It's a notch below full MCP fidelity, but it works in any agent that can browse the web.

You can also grab the same markdown yourself with Copy as markdown in a review's share section, or by appending /markdown to the review URL.

How CobaltCapture fits in

You capture findings — drag a box around each broken region, dictate the context, publish. The result is one link. The MCP server and the markdown export both read from that same review, so whichever path your agent takes, it gets the screenshots and your intent in a form it can act on. See the full reference at for coding agents.

Frequently asked questions

What is the Cobalt MCP server?

A remote Model Context Protocol endpoint at https://cobaltcapture.com/mcp. Add it to an MCP-capable agent (Claude Code, Cursor, Codex) and it exposes one tool, get_review, that returns a review's notes as text plus each annotated screenshot as a real image.

How do I add it to Claude Code?

Run claude mcp add --transport http cobalt https://cobaltcapture.com/mcp. Then ask your agent to read a review by its URL and it will call get_review, receiving the screenshots as image blocks it can actually see.

Do I need an account or API key?

No. The MCP server serves public reviews with no auth — the same access as opening the review link in a browser. There is nothing to sign up for and no key to configure.

What if my agent doesn't support MCP?

Just paste the review link. When an AI fetcher opens a Cobalt review URL it gets the markdown export — the annotations described in text plus a note telling the agent to fetch and view each screenshot. MCP gets you full image fidelity; the link works everywhere.

Why does the image matter — isn't the text enough?

A screenshot with an arrow pointing at the exact broken element carries information the words around it can't. The MCP server hands the agent the composed image with the annotation burned in, so it sees precisely what was marked instead of guessing from a description.

Frequently asked questions

What is the Cobalt MCP server?

A remote Model Context Protocol endpoint at https://cobaltcapture.com/mcp. Add it to an MCP-capable agent (Claude Code, Cursor, Codex) and it exposes one tool, get_review, that returns a review's notes as text plus each annotated screenshot as a real image.

How do I add it to Claude Code?

Run: claude mcp add --transport http cobalt https://cobaltcapture.com/mcp. Then ask your agent to read a review by its URL and it will call get_review, receiving the screenshots as image blocks it can actually see.

Do I need an account or API key?

No. The MCP server serves public reviews with no auth — the same access as opening the review link in a browser. There is nothing to sign up for and no key to configure.

What if my agent doesn't support MCP?

Just paste the review link. When an AI fetcher opens a Cobalt review URL it gets the markdown export — the annotations described in text plus a note telling the agent to fetch and view each screenshot. MCP gets you full image fidelity; the link works everywhere.

Why does the image matter — isn't the text enough?

A screenshot with an arrow pointing at the exact broken element carries information the words around it can't. The MCP server hands the agent the composed image with the annotation burned in, so it sees precisely what was marked instead of guessing from a description.