What is Iris?
Iris is a small Rust program that drives an installed Chrome, Chromium, Edge or Brave over the DevTools Protocol and saves screenshots of live pages. It waits for fonts, images and entrance animations, scrolls full pages to trigger lazy loading, and can frame the first element matching a CSS selector with padding. Defaults are a 1440x900 viewport at 2x; presets cover iPhone and iPad, and output is PNG, JPEG or WebP.
The AI angle is explicit: the same binary runs a local stdio MCP server with `iris mcp`, exposing one tool, capture, that returns the image inline to the agent with its dimensions. The README gives a one-line setup for Codex and a JSON block for any other MCP client, and a `--json` flag on the CLI emits machine-readable results for scripts.
The repository is MIT-licensed, maintained by one developer, has about 540 GitHub stars and 29 forks, and is at v0.4.1 (18 August 2026), published on crates.io as iris-screenshot.
Is Iris worth it?
A coding agent that edits a front end usually cannot see what it changed. Iris fixes that narrowly: the agent calls capture with a URL, an optional selector and padding, and gets pixels back in the same tool result rather than a file path it then has to open. Bare localhost addresses default to plain HTTP, so pointing it at a dev server on port 3000 works without fiddling.
The author publishes reference timings rather than adjectives: on an Apple M2 Max, one CLI capture took a 1.00 s median, the first MCP capture 965 ms and later ones a 366 ms median, because the MCP server keeps one Chrome process alive. Those numbers come from a local test page and the README says they are not a guarantee. Full pages taller than Chrome's roughly 16,000 pixel limit fall back to 1x scale, and the report says so.
The limits are deliberate. There is no clicking, typing, login or scripting, cross-origin iframes are not captured, only the first selector match is taken, and each launch uses a fresh browser profile, so pages behind your own sign-in are out of reach. The capture tool's optional output path is written wherever the server process can write, so keep the server's working directory sensible. Development has been quiet since early September apart from dependency updates.
Next to Browser Use or Stagehand, which let an agent operate a site, Iris only looks. That is the point: one tool with one argument shape gives an agent less to get wrong and gives you less to secure.
Where it earns its place
A front-end developer using Codex on a React app
Let the agent see whether the pricing card it just restyled looks right
Register Iris with codex mcp add iris -- iris mcp, then ask the agent to capture localhost:3000 with the selector #pricing-card and 24 px padding after each change. The image comes back inline; nothing is written to disk unless an output path is given.
A designer at a small studio maintaining client sites
Take before-and-after screenshots of many pages in light and dark mode
Pipe a list of URLs into iris - -o shots/ with --dark for the second pass; captures run concurrently in one browser and a failed page is marked without stopping the batch. No account or subscription is involved.
Before you commit
- Setup
- Install with the curl script or cargo install iris-screenshot (Rust 1.88+ to build), make sure a Chrome-family browser is installed, then register iris mcp in your MCP client; a fresh agent task may be needed before the capture tool appears.
- Integrations
- Any MCP client over stdio, with a documented one-liner for Codex; the CLI's --json output suits shell scripts and CI. Chrome, Chromium, Edge or Brave as the renderer.
- Data and privacy
- Runs entirely on your machine with a fresh browser profile per launch. Images go to the agent that called the tool, and through it to that agent's model provider. No telemetry is mentioned in the README.
- Limits
- One image per call; first selector match only; no cross-origin iframe contents; no interaction or authentication; full pages over about 16,000 pixels drop to 1x; 30 second default timeout per page.
What’s good
- MCP server built into the same binary, one tool only
- Waits for fonts, animations and lazy-loaded content
- Element capture by CSS selector with padding
- Published, reproducible timing method
What’s not
- No interaction, login or scripting by design
- Fresh profile each launch, so no signed-in pages
- Single maintainer; little activity since September
- Needs a local Chrome-family browser
Best for
Developers who want Claude Code, Codex or another MCP-capable agent to check its own front-end changes visually without wiring up a browser-automation framework.
Iris alternatives
Iris: common questions
- Is Iris worth it?
- Iris does one job for coding agents, taking a trustworthy picture of a page, and does it with less setup and less surface than a full browser-automation stack. We rate it 3.9 out of 5.
- How much does Iris cost?
- Free and open source (MIT). Needs only a Chrome-family browser already installed.
- What are the downsides of Iris?
- No interaction, login or scripting by design; fresh profile each launch, so no signed-in pages; single maintainer; little activity since September; needs a local Chrome-family browser.
- Who is Iris best for?
- Developers who want Claude Code, Codex or another MCP-capable agent to check its own front-end changes visually without wiring up a browser-automation framework.
- Is Iris safe for company data?
- Runs entirely on your machine with a fresh browser profile per launch. Images go to the agent that called the tool, and through it to that agent's model provider. No telemetry is mentioned in the README.
- What are the best Iris alternatives?
- The closest alternatives we have reviewed are Browser Use, Stagehand and Lightpanda.
Answers checked Nov 18, 2026.