Introduction
Most browser automation breaks the same way: a selector drifts, a modal appears, a button renames itself, and your overnight job dies at 3 AM. The usual fix is babysitting — patching selectors, adding sleeps, wrapping everything in try/catch. Browser Harness takes a different angle. It's a thin layer on top of the Chrome DevTools Protocol where the agent rewrites its own helpers when they fail, mid-task, without you in the loop.
That's the pitch. I spent enough time in the repo and driving it against real flows to have an opinion. The short version: the core idea is good, the implementation is genuinely lean, and the project is too young to bet a production pipeline on. Here's the long version.
Key Features
Self-Healing Agent Loop
This is the differentiator. When a helper function fails — selector not found, click misses, navigation times out — the agent doesn't just retry. It inspects the page state, rewrites the helper, and continues the task. In practice this means a script that breaks every other week on a third-party site is more likely to just keep working. It's not magic; it still fails on hard cases. But the failure rate on flaky-but-recoverable tasks drops noticeably.
Direct CDP Integration
No Playwright, no Selenium, no wrapper-of-a-wrapper. Browser Harness talks directly to Chrome DevTools Protocol. The latency difference is real if you're chaining hundreds of operations per task. It also means you can drop down to raw CDP calls when you need to — useful for network interception, request modification, or anything the higher-level abstractions hide from you.
Three Deployment Modes
- Local Chrome — point it at your own browser. Best for development, debugging, and anything where you want to watch what's happening.
- Stealth Browsers — managed fleet with anti-detection headers. For when Cloudflare and friends keep flagging your scraper.
- Browser Use Box — dedicated 24/7 sandbox with persistent sessions. For agents that need to stay logged in and run continuously.
MIT License on the Core
The harness itself is open source. You can self-host it, fork it, audit it. The paid tiers are infrastructure (stealth fleet, persistent boxes), not the engine.
Pricing Breakdown
| Plan | Price | What You Get |
|---|---|---|
| Open Source | Free | MIT-licensed harness, self-hosting on local Chrome, full CDP access, community support |
| Stealth Browsers | Custom | Managed stealth browser fleet, anti-detection headers, no local Chrome required |
| Browser Use Box | Custom | 24/7 always-on execution, dedicated sandbox, persistent agent sessions |
The honest read: if you're a solo developer or a small team, the open source tier is everything you need to evaluate the tool and build real things. The custom-priced tiers are where things get opaque. There's no public price list for stealth browsers or Browser Use Box. You'll need to talk to them, and if you've been around long enough you know what "contact us" usually means — it depends on your volume, and the answer won't be cheap.
Pros
- MIT licensed core. No vendor lock-in on the actual harness. If they disappear tomorrow, you keep what you built.
- Self-healing actually reduces babysitting. The mid-task helper rewriting isn't a gimmick — it cuts the class of failures that usually need a human to look at them.
- Thin CDP layer. Lower overhead than Playwright or Selenium wrappers, and you can drop down to raw protocol calls when you need precision.
- Deployment flexibility. Local for dev, stealth for hostile sites, dedicated box for always-on agents. You're not forced into one model.
Cons
- Opaque pricing on the paid tiers. No public numbers for stealth or Browser Use Box. Hard to evaluate without a sales conversation.
- Not a no-code tool. You need to be comfortable wiring up an LLM agent loop. If you're looking for a point-and-click scraper builder, this isn't it.
- Very new project. Limited community, limited production track record, limited Stack Overflow answers when something breaks at 2 AM. You're an early adopter, with all that implies.
- Chrome only. CDP means no Firefox, no Safari, no WebKit. If your testing matrix includes those, look elsewhere.
Who Is It For
Browser Harness is for developers who are already building browser agents and are tired of two things: their automation breaking on selector drift, and their existing tooling feeling like a wrapper around a wrapper. If you've used Browser Use and wanted more raw control, or you've used Playwright and wanted less ceremony, this sits in the gap.
It is not for: teams that need production SLAs today, anyone who needs cross-browser coverage, anyone evaluating tools by counting GitHub stars or community size, or anyone who wants a hosted no-code workflow builder. Look at Firecrawl for managed scraping or E2B for sandboxed code execution if those are closer to your need.
The sweet spot is a developer or small team building an internal agent that touches a handful of sites, where the self-healing loop earns its keep by not waking you up at night, and the open source license means you can fork it if the project stalls.
Verdict
Browser Harness gets a 6.5/10. The self-healing idea is genuinely useful and the CDP-direct design is the right call for the audience it's aimed at. The MIT core means the downside of trying it is bounded — worst case, you learn something and move on with your code intact.
The reason it isn't higher: it's too new. Documentation is thin. The community is small. The opaque pricing on the cloud tiers makes it hard to plan around. If you ship production automation today and an outage costs you money, you should not be the first team in your industry running this in prod.
Recommendation: If you're building browser agents and the self-healing pitch resonates, clone the repo, run it locally against one of your flakier flows, and see if the failure rate drops. That experiment costs you an afternoon. If it works, keep it on the side until the project matures — six to twelve months from now it could be the obvious choice for this layer of the stack. If you need something rock-solid right now, stick with what you have and check back.