Annotated screenshots for AI agents,
not directions in words.

Words versus a picture

The same question, answered two ways.

Without AInotate

Where do I change the text size on Wikipedia?

Look at the top right of the article. There is an Appearance menu, usually next to the Tools menu, though on some layouts it sits in the sidebar instead. Inside it, under Text, pick Standard, Small or Large. If you don't see it, try widening the window or check the hamburger menu.

With AInotate

Where do I change the text size on Wikipedia?

Open Appearance and pick 1 for bigger text. 2 and 3 set full width and dark mode.

What people use it for

“Where is it?”

Settings, menus and buttons hidden behind icons. The agent answers with the screen itself, the control boxed and numbered.

Bug reports and tickets

What is wrong in red, what is right in green. Web shots made with shoot black out emails and tokens before you see them.

Guides and changelogs

One image per step, exported as a Markdown, HTML or PDF guide. Before and after plates when the UI changes.

“Your turn”

Permission switches, 2FA codes, payments: the agent does not click for you. It shows what to press and says what the click does.

Labels stay off the content. AInotate maps the text and icons on the page and puts each label in free space near its target, with an arrow that curves around what it would cover. Left: a fixed offset. Right: AInotate.

Install

The plugins bring two parts together: the skill, which tells the agent when and how to annotate, and the MCP server, which does the capturing and drawing. The plugin installs need uv.

{"mcpServers": {"ainotate": {"command": "uvx", "args": ["--from", "ainotate[all]", "ainotate", "mcp"]}}}

With the command line tool installed, {"command": "ainotate", "args": ["mcp"]} works too, and ainotate install-skill adds the skill for Claude Code and Codex. Config paths for every client: MCP setup.

Get started

1. Step 1: Check the machine

   Ask your agent to run the doctor tool, or run ainotate doctor. It checks screen capture permission, the browser and OCR, and prints the exact fix for anything missing, such as the command that downloads Chromium for web capture (python -m playwright install chromium).

2. Step 2: Ask your agent

   Try “Show me where to switch Wikipedia to dark mode.” The agent captures the page, finds the control, draws a numbered step, looks at the result and sends you the image. With the skill installed it also does this on its own whenever an answer depends on where something is on the screen.

3. Step 3: Or drive it yourself

   A shot is a small JSON spec: where to capture and what to mark. Save it as hn.json and run ainotate shoot hn.json --draft.

   {"url": "https://news.ycombinator.com",
    "frame": {"chrome": "browser"},
    "marks": [
     {"type": "step", "n": 1, "target": {"text": "new", "exact": true}, "label": "Newest"},
     {"type": "step", "n": 2, "target": {"text": "login"}, "label": "Sign in"}]}

   Mark types: step, box, arrow, click, highlight, spotlight, magnify, keys, text and redact; the first four take a label. blur and pixelate only tone down clutter and are never used for secrets. Every command, the full spec, the Python API and exit codes are in the README.

Updating: brew upgrade ainotate, pipx upgrade ainotate, /plugin marketplace update ainotate in Claude Code, or the newer .mcpb for Claude Desktop. Stuck? Open an issue on GitHub with the output of doctor.

Privacy

DontPayFull never receives your screenshots. AInotate runs on your machine and hands its images to the agent you use. If that agent is Claude, Codex, Gemini or Cursor, the image becomes part of the conversation and goes to that provider under their terms.

• No account, no telemetry, no server of ours. The only network traffic is loading the pages you capture and the downloads you install.
• Web shots made with shoot black out emails, phone numbers, cards, API keys and tokens. For other images, turn it on with --privacy auto.
• Solid blocks only, like the ones beside this list. Blur can be reversed, so it is never used to hide data.
• Detection is best effort and raw captures are not redacted. The agent checks a preview before it sends the final image; look at it too.

Full privacy notes