An agentic browser harness: an AI side panel & orb in your own logged-in Chrome, Edge, Arc, Brave or even Firefox run by the agent/subscription you already use (Claude Code, Codex, Antigravity). Free, open source, no API key. Also an MCP server.
First run
A tour of the side panel, three instructions to try first, and what to check if the agent does not answer. It assumes you have installed and paired.
Open the side panel
Click the Browsentic toolbar icon, or open it from Chrome's side panel menu. Clicking the icon again closes the panel. In Firefox the icon opens and closes the Browsentic sidebar.
The panel, top to bottom
| Status pill | Connection state. Click it to open the pairing form and the agent switcher. |
| Sessions strip | One row per tab that has a conversation: its live title, a pulsing dot while it is working, how many messages it holds. Collapses to a single line. See Conversations. |
| Chat · History · Skills · Recordings · Schedules | The five sections. Chat is where you talk; the others are covered in Features. |
| Timeline | Every action as it happens, with what it targeted and what came back. |
| Composer | Type, or press the mic to dictate. |
Your first instruction
Open any ordinary website (not a chrome:// page, which cannot host a content script) and try:
what's on this page?
The agent calls page_getPageInfo, then answers. That round trip shows all four pieces working:
panel → Browsentic Bridge → agent CLI → browser.
Then try something that acts:
scroll to the bottom
That one carries a ⚡ and returns at once: it matched the local command grammar and ran in the browser, with no agent involved. See Instant commands.
Then something multi-step:
find the search box, type "wireless headphones" and show me what comes back
Follow-ups continue the same conversation, so "now click the second one" works.
What to expect the first time something is consequential
Submitting a form pauses the run and asks you, with Allow, Deny and Always on ‹host›. The last grants that one action on that one site and stops asking.
Denying is final: the agent is told to report it and stop, not to find another route to the same effect. Approvals covers the whole gate.
If it does not answer
Check in this order:
browsentic status: is your browserconnected?browsentic agent: is the agent CLI installed and ready?browsentic logs: run starts, routed skills, every tool call
Troubleshooting maps symptoms to fixes.
Where to go next
- Features: one page per capability
- Choosing an agent: to run one of the other supported agent CLIs
- Configuration:
~/.browsentic/config.json - Limits: where Browsentic stops, worth reading early