On this page
If you want an agent-browser alternative because your coding agent can drive the page but still cannot tell whether the change worked, Reticle adds a verdict from network, app state and console and a replay with no model. If you want the fastest, cheapest way for an agent to click around any site, Vercel's agent-browser is an excellent driver and needs no setup.
What agent-browser does
Vercel Labs describes it as a "Browser automation CLI for AI agents. Fast native Rust CLI." (GitHub).
The agent runs snapshot to get an accessibility tree with short refs such as @e1, then acts by ref, for example click @e2. Compact refs keep the agent's context small. It ships a Claude plugin and skill and an agent-browser mcp stdio server. It is free and open source under Apache-2.0, with about 44,500 GitHub stars on 2026-10-25, and v0.39.2 shipped on 2026-10-22. It drives; it does not assert, map to source, read app state or replay.
What Reticle does
Reticle checks an AI coding agent's change from inside the running web app and returns a verdict with evidence. It reads the network calls, the application state and the console, and on React and Svelte it names the file:line to fix.
Your agent calls it through MCP. Flows it drives are saved and replay deterministically, with no model in the loop. It also works on sites you don't own: it drives the page in its own browser and injects its SDK, reading network, console, DOM and routing; app state and file:line need the SDK in your own build. The SDK is open source (Apache-2.0).
agent-browser vs Reticle at a glance
| agent-browser | Reticle | |
|---|---|---|
| Job | Drive a browser cheaply | Decide whether a change worked |
| What the agent gets | Accessibility snapshot with refs | Pass, fail or cannot tell, with evidence |
| Assertions | None built in | Expected outcome per step: request count, state, console, DOM |
| App state and source | Not read | Store state and file:line (with the SDK in your build) |
| Re-running | The agent drives again | Deterministic replay, 0% flake over 8 replays |
| Setup | None | MCP config plus a small SDK, or one line to your agent |
| Licence | Apache-2.0 | SDK Apache-2.0; server FSL-1.1 |
Where agent-browser is the better choice
- Zero setup on any site. Install the CLI and the agent drives any URL.
- Token-cheap driving. Short refs keep snapshots small, which matters on long sessions.
- General browsing tasks. Research, form filling and scraping are its home ground, not Reticle's.
- Adoption. A very large user base means lots of examples and answers.
Where Reticle is the better choice
- You need a pass or fail, not a description. Reticle declares what a step should cause, then checks the request count, the state and the console after it settles. A double-submit or a write that never persists fails even when the page looks right.
- You want a regression check. A saved flow replays with no model: about 480 tokens to re-verify four flows, about 251 times less than a model re-driving them.
- You want the fix located. On React and Svelte the verdict names the source line.
On 88 seeded bugs, Reticle caught 85 with no false alarms (scorecard).
Can you use both?
Yes. They do different jobs: drive with agent-browser, assert with Reticle. The same split applies to the other drivers on the MCP browser testing page and in our Playwright MCP comparison.
Try Reticle
Paste one line to your coding agent, and it installs and wires the SDK in your app:
Follow https://raw.githubusercontent.com/reticlehq/reticle/main/SKILL.md
FAQ
What is Vercel agent-browser?
Vercel agent-browser is an open-source Rust CLI that lets AI coding agents drive a browser. The agent takes an accessibility snapshot with short element refs and acts on them, which keeps context small. It also runs as an MCP server and ships a Claude skill. It drives the page but does not decide whether a change worked.
What is an alternative to agent-browser for testing?
An alternative to agent-browser for testing is Reticle, which adds the part a driver leaves out: a pass or fail decided from network, app state and console, and a deterministic replay with no model. Playwright MCP and Chrome DevTools MCP are other drivers; they also leave the verdict to the agent.
Is agent-browser better than Playwright MCP?
agent-browser and Playwright MCP are both ways for an agent to drive a browser, and the better one depends on your setup. agent-browser is a native CLI with compact refs; Playwright MCP is Microsoft's MCP server with cross-browser support. Neither decides whether your change worked, which is the job Reticle adds on top.
Why Reticle
Your agent says it works. Reticle checks.
- Checks each change inside the running app: network, state and console, not a screenshot.
- Returns a pass or fail with the evidence and the file:line to fix.
- Open-source SDK on your machine, or the platform on every preview and PR.
No credit card required.