The agent drives the browser; you keep a hand on the wheel
Browser agents fall apart the moment a site needs a human — this one shares a single live Playwright session, so takeover is just reaching for the mouse.

What it does — Auto Browser is an MCP-native control plane that hands LLM agents a real Chromium browser driven by Playwright. Agents get screenshots, DOM summaries, OCR excerpts, tab control, downloads, and network inspection over MCP (HTTP or a bundled stdio bridge) or plain REST. The same live session is mirrored through noVNC, so a human operator can watch — or seize the mouse mid-run — without any handoff ceremony. Named auth profiles let a person log in once, then agents reopen fresh sessions that are already signed in.
The interesting bit — Takeover isn’t a replay or a screenshot review; it’s the identical browser instance, so the agent’s context survives the human’s intervention and vice versa. The stranger feature is the convergence harness: it records tamper-checked traces of successful runs, verifies completion, and stages them as skill candidates with signed provenance — promotion stays explicitly human-reviewed. The pitch is agents that prove they can repeat themselves correctly, not just act once.
Key highlights
- MCP over HTTP plus a bundled stdio bridge; works with Claude Desktop, Cursor, or curl-first REST callers
- Named auth profiles: a human logs in once, agents reuse the signed-in state in new sessions
- Safety rails as product surface — approval gates, operator identity, PII scrubbing, Witness receipts, and
strict/balancedpolicy presets - Local-first: the full stack (Chromium, Xvfb, x11vnc, noVNC, FastAPI controller) runs on your own box, with published ports bound to loopback by default; a Codespaces demo spins up in about 90 seconds
- Refreshingly explicit non-goals: no CAPTCHA solving, no unauthorized scraping, no identity-spoofing tooling
Caveats
- The compliance presets were renamed from HIPAA/SOC2/GDPR/PCI-DSS to generic
strict/balanced(legacy names still work as deprecated aliases) — read the hardening docs before leaning on the compliance vocabulary - The default
curatedMCP tool profile is deliberately compact; the convergence harness and the full tool surface sit behind the non-defaultfullprofile - Recent release notes mention stripping unwired social/Veo3 modules — a hint of scope churn in a project already carrying a lot of surface
Verdict — Worth a look if your agent workflows hit brittle, login-walled, occasionally-needs-a-human sites; internal dashboards and admin tools are the stated sweet spot. Skip it if you wanted a lightweight headless scraping library or anything in the explicitly rejected CAPTCHA-and-scraping category.
Frequently asked
- What is LvcidPsyche/auto-browser?
- Browser agents fall apart the moment a site needs a human — this one shares a single live Playwright session, so takeover is just reaching for the mouse.
- Is auto-browser open source?
- Yes — LvcidPsyche/auto-browser is open source, released under the MIT license.
- What language is auto-browser written in?
- LvcidPsyche/auto-browser is primarily written in Python.
- How popular is auto-browser?
- LvcidPsyche/auto-browser has 914 stars on GitHub and is currently holding steady.
- Where can I find auto-browser?
- LvcidPsyche/auto-browser is on GitHub at https://github.com/LvcidPsyche/auto-browser.