Local-first co-browse runtime · open source

An AI agent inside your real, logged-in browser.

Browgent is not another chat sidebar. It is a desktop runtime where an agent acts on your actual tabs and session — and you are always one click from taking the wheel. Attach Claude Code (MCP) or Playwright (CDP) to the same tabs, behind a policy engine with exportable trajectories.

Download for macOS

or watch the 60-second flow ↓  ·  Star on GitHub  ·  Builders guide  ·  Become a design partner

Proof you can run today: clone it, npm run dev, attach Claude Code over MCP, and take over a login mid-task — no signup, no cloud, no waitlist. MIT-licensed, local-first, model-agnostic. This is an early runtime, honest about its stage: the wedge below is the pitch, not a user chart.

Local-first cookies Claude Code / Cursor MCP Playwright CDP Takeover → resume Policy + trajectory audit

The flow: agent acts → hits a login wall → you take over → agent resumes → policy confirms

app.yourcompany.com/dashboard ◆ AGENT
Run monthly report
1 Agent acts on your real, logged-in tab
login.yourcompany.com/sso ❚❚ AGENT PAUSED
Single sign-on required Agent won't guess credentials — it asks you.
2 Hits an SSO / login wall — agent stops and calls ask_human
login.yourcompany.com/sso ● YOU
You take the wheel — on the same tab you@yourcompany.com ••••••••••• ✓ signed in
3 Human takeover — you log in; the session stays local
app.yourcompany.com/dashboard ◆ AGENT
Session intact — resuming the task
4 Agent resumes on the same tab — no lost context, no re-login
app.yourcompany.com/checkout ◆ AGENT
5 Policy gate pauses the risky step until you confirm

Looping illustration of the co-browse + takeover flow — not a screen recording. A real demo drops in above.

True co-browse

Same tab tree. Ownership badges. Takeover → resume without losing the session.

Builder attach

STDIO MCP for Claude Code / Cursor. Playwright over CDP on the same cookies.

Policy + audit

Confirm sensitive actions, host gates, modes, exportable trajectory JSON for evals.

Not these

BrowserOS is a browser for end-users to automate their own browsing; Browgent is a runtime you attach agents and Playwright to, with a human in the loop. Comet, Dia, and Atlas are the consumer-category reference — polished AI browsers with a closed assistant.

ProductJob
BrowgentLocal co-browse runtime you attach agents + Playwright to (HITL)
BrowserOSOpen-source agentic browser for end-users (ties us on OSS + local)
Comet / Dia / AtlasClosed consumer AI browsers with a built-in assistant
Browserbase / SteelCloud multi-tenant headless fleets
Playwright MCP aloneDev tools without a shared human UI or policy chrome

60-second builder path

  1. Install or npm run dev
  2. Status bar: mcp · :17342 (token in userData)
  3. npm run mcp → wire Claude Code
  4. In-app: Run demo or a recipe chip (takeover leads)
  5. Take over on a login wall → Resume → Export trajectory

Who this is for

Beachhead: developers building browser agents that keep failing on real login / SSO / CAPTCHA walls. Browgent gives those agents a local, logged-in session and a human who can step in and hand control back — with policy and an audit trail.

Also useful for HITL automation/ops teams and eval/safety labs that need exportable trajectories.

Design partners

Building browser agents that die on login walls? We're onboarding a small set of partners for free support in exchange for honest feedback (and an optional quote).

Apply as design partner