A real, full-GUI browser your AI agent can actually drive — log in with credentials it never sees, complete checkouts, render live JavaScript, and record every session. Ships as an MCP server, Claude extension, GPT plugin, and Claude Code skill.
General agent browsers refuse the useful things. AgentBrowser replaces the refusal with architecture — origin-bound secrets, full recording, and real auditability.
Click, type, select, hover, scroll, press keys, run JS — with DOM-measured actionability, so no flaky pixel-guessing.
An origin-bound vault injects username, password, and TOTP at fill time. The agent triggers a login and never receives the secret.
It does the things other agent browsers refuse — sign-ins, 2FA, multi-step purchases — under explicit, auditable permission.
Full-session MP4 capture of every run, so you can replay exactly what your agent saw and did.
A durable, replayable event log — HAR for network, incremental DOM, every console line — queryable at any timestamp.
Live tail over WebSocket / SSE / NDJSON. Watch an agent work in real time, resume from any point, gaps self-heal.
On-node watchers fire the instant a condition is met — catching an 800 ms toast that a polling agent misses entirely.
Real Chrome, real hydration — see the page your users see, not a static fetch of the HTML.
Route through the egress of your choice; test at real device resolutions with true multi-touch that CDP alone can't deliver.
Connect over MCP (or the Claude/GPT integrations), and every instance links back to the AgentBrowser control plane for identity, permissions, and the credential vault.
Add AgentBrowser as an MCP server, Claude extension, GPT plugin, or Claude Code skill. Your agent gets tools to navigate, act, snapshot, record, and log in.
Register a login in the vault with an origin allowlist. Your agent requests "log in as test-user" — the secret is injected into the page, never returned to the model.
The agent drives a real browser: forms, checkouts, multi-step flows across desktop and mobile. Stream it live or review the recorded bundle afterward.
Every action, network call, console line, and vault access is logged. Approve sensitive vault requests from the web, desktop, or mobile app.
The same browser, exposed the way your stack already works.
Either way, instances connect to the AgentBrowser control plane for control, identity, and vault access.
Your box, your server, your laptop.
Managed instances, on demand.
Sign up and test for free. When you scale, usage is metered on real infrastructure — not per-token — so the AI vision calls stay BYOK.
Join the early access. We'll set you up with a workspace, an API key, and free instance-time to test.