Visual feedback for AI coding agents.

Annotate any screenshot in your browser. Click to drop pins, drag to draw boxes, type comments. The agent picks up structured coordinates and your comment, and works each annotation as a discrete fix.

curl -fsSL https://pinpoint.maferland.com/install.sh | bash

/plugin marketplace add maferland/pinpoint
/plugin install pinpoint@pinpoint-marketplace
/pinpoint:install

How it works

Your agent drives. You point at what's wrong.

The agent takes a screenshot Of whatever it's building. A page in your dev server, a Storybook story, an iOS simulator — anywhere with pixels.

You pin what's wrong Click to drop a pin, drag to draw a region, type your comment. Hit Send when you're done.

The agent fixes what you pointed at Each pin becomes a task. The agent works through them and asks for another round of pins when it's ready.

What Pinpoint sends back

{
  "annotations": [
    {
      "number": 1,
      "image": "/tmp/screenshot.png",
      "box": { "x": 10.2, "y": 5.3, "width": 35.0, "height": 12.5 },
      "comment": "Button text is truncated on mobile"
    }
  ]
}

Use it with your agent

Pinpoint speaks two protocols. Pick whichever fits where you already work.

Claude Code

Slash command, no setup beyond install.

/pinpoint:review /tmp/screenshot.png

Claude Code via MCP

The install registers an MCP server so Claude can drive Pinpoint inline — no slash command required.

Direct CLI

Spawns the server, opens the browser, blocks until you hit Send, prints JSON to stdout. Pipe it anywhere.

pinpoint review screenshot.png \
  --context "Login page after auth changes"

Try the loop in 10 seconds

After install, no screenshot needed. Run from Claude Code (or your terminal) and a sample session opens in your browser with starter pins.

/pinpoint:demo