Agent-first · works over MCP

Your agent ships it. You just review.

Flip the loop around: your coding agent publishes its own work and hands you the review. It captures the build, gives you a link, waits for your flags, then pulls them back as structured JSON and revises — all over MCP. You do the one part agents can't: deciding.

Get your account Connect your agent

Free to try — your agent can run a full review round on the free tier.

The loop

Six calls, one review round.

comment.gg gives your agent a small set of MCP tools that map exactly onto the review loop. It drives the whole thing and only ever waits on you for the human judgment call.

  1. Connect

    Add comment.gg to your agent once, over MCP. One endpoint, OAuth sign-in — no API keys to paste around.

  2. Capture & share

    Your agent calls create_review and uploads its real dist/ — not a screenshot — then hands you the review link.

  3. Wait for you

    It parks on await_feedback and watches the review live. You open it whenever you want.

  4. Pull the flags

    You drop flags like any reviewer. The agent reads them with get_feedback — structured, with selector, text quote, and screenshot.

  5. Revise

    It fixes, then add_version pushes the next round into the same review. Your open flags carry forward automatically.

  6. GG

    When you're happy, the round closes. Nothing left open, nothing lost between iterations. Go next.

Freemium · connect in a minute

Point your agent here.

comment.gg runs a remote MCP server. Add the endpoint to any MCP-capable agent, sign in once, and it can create reviews, push versions, and read your flags.

https://mcp.comment.gg/mcp

Claude Code

One command wires it up. Your agent signs in through your comment.gg account on first use.

Already building with comment.gg in your team's flow? The same review link works for humans and agents — bring people in too.

$ claude mcp add --transport http \
    comment-gg https://mcp.comment.gg/mcp
// any MCP client — .mcp.json
{
  "mcpServers": {
    "comment-gg": {
      "url": "https://mcp.comment.gg/mcp"
    }
  }
}
Built for the agent that built it

Structured feedback, not a paste of vibes.

Every flag comes back as JSON the agent can act on — the exact selector, the text it quoted, a screenshot, and your change request, verbatim. No re-typing what you meant. And when an agent is in the review, it shows up as itself — a flag marked “· agent,” never disguised as a human.

Get started free 
{
  "status": "open", "page": "/pricing",
  "raisedBy": "Peter",
  "target": {
    "selector": "#plans .tier:nth-child(2) h3",
    "textQuote": "Pro — $12/mo",
    "screenshot": "https://…/flag-8a1c.png"
  },
  "request": "Change Pro to $14/mo, add 'billed annually'."
}
GG.

Let the agent close the loop.

You shouldn't be the one taking screenshots and writing tickets back to your own agent. Hand it the review layer and just decide. GG — go next.

Get started free