imge is in early preview. Set it up withnpx @imge/cli setup
Open SourceLocal-firstMIT

See what your agent saw.

Session replay for browser agents, on your machine.

npx @imge/cli setup
Then try it

Using imge browser, go to example.com and complete add to cart flow.

localhost:4747
Checkout flowStep 4 of 9
ScreenshotAgent view
What was on screen
everyday objectsShop   About   Cart (0)
Desk essentials
Everyday notebook
A simple place for your ideas.
You may also like
What the agent read
- main:
- heading "Everyday notebook"
- paragraph "A simple place for your ideas."
- button "Add to cart" [ref=e12]
- button "Add to cart" [ref=e31]
- button [ref=e40]
- navigation: Shop, About, Cart (0)
click button "Add to cart" (2 matches)
Same name: 2 elements
4 / 9
Agent says done.Cart is empty.You have no idea why.

How it works

One setup. Every step in view.

01

Add one line to your MCP config

npx @imge/cli setup

02

Your agent browses as usual

Claude Code, Codex, Cursor, and any MCP client.

03

Replay every step it saw

Screenshot and agent view, side by side. Same-name warnings. Sessions stay on disk; password and card fields are masked.

Open the local player with npx @imge/cli at localhost:4747.

FAQ

A few good questions.

Which agents does it work with?

Any MCP client using Playwright MCP: Claude Code, Codex, Cursor, Claude Desktop and VS Code. Cloud agents like ChatGPT Agent are not covered yet.

Does any data leave my machine?

No. Sessions are files in a local folder, the CLI sends no telemetry, and password and card fields are masked.

Does it change what my agent does?

No. It passes calls through to Playwright MCP and only records what comes back.

Playwright already has tracing. Why imge?

Playwright's tracing is great, and imge doesn't replace it. They answer different questions. Trace Viewer is built for debugging tests: DOM, network and console. imge is built for checking an agent's work: it shows each screenshot next to the text the agent actually read, and flags things that confuse agents, like two elements with the same name. Use Trace Viewer when a test fails, and imge when your agent says "done" and you want to see what really happened.

Does it replace Langfuse, LangSmith or Laminar?

No. Those tools show model calls, tokens and cost for teams building agents. imge shows the page your agent was on. You can run both.

Is it free?

Yes. It's MIT licensed and open source, with no account.

Get started

Replay your first agent session in a minute.

Free and open source. No account. Runs on your machine.

npx @imge/cli setup