What is Browsentic?
Browsentic puts an AI side panel in Chrome, Edge, Brave, Arc, Vivaldi, Opera or Firefox and runs it on an agent CLI you already have, such as Claude Code, Codex or Antigravity, with six more marked beta. There is no API key or account of its own. The extension connects over a local WebSocket to Browsentic Bridge, a Node daemon on 127.0.0.1, which spawns the agent CLI and hands it 52 page tools. The same bridge is also an MCP server over stdio, so Cursor, Zed or Claude Code can drive the browser directly.
Because it works in your real profile, the agent sees every site you are logged in to. A policy layer asks before form submissions, uploads, downloads, off-site navigation and captcha answers, refuses some actions outright, and tries to seal credentials before page text reaches the model.
The repository is Apache-2.0 (MIT up to 0.7.15), written in TypeScript, and has about 300 GitHub stars. It was created in July 2026 and is maintained by one developer; v0.9.2 was released on 9 November 2026.
Is Browsentic worth it?
Most browser agents work in a fresh, logged-out browser, which is safe and often useless: the jobs people want done sit behind a login. Browsentic takes the opposite route. It drives the browser you are looking at, with your cookies and sessions, from a side panel powered by the Claude Code or Codex subscription you already pay for. That makes it cheap to try and much riskier than a sandboxed browser, and the project says so on nearly every page.
The guardrails are more thorough than most. Reading and clicking need no approval; submitting a form, attaching a file, letting a page download, leaving the sites a run is about and answering a captcha all stop for a yes. Raw HTML reads and network bodies are refused, MCP clients cannot answer prompts so their gated actions are denied by default, and the environment passed to the spawned CLI is stripped of cloud keys and tokens. How tightly the CLI itself is fenced depends on the agent: the limits page rates Claude Code's per-run tool allowlist as the strongest and says Codex can still read any file you can.
The security page is candid about what is left. Pairing controls which browser connects, not which local program: anything running as your user can read `~/.browsentic/daemon.json` and drive a paired browser. Prompt injection from a hostile page is mitigated, not solved. Credential sealing is a pattern filter, so a short passphrase in plain prose can slip through, and held secrets sit unencrypted in the extension's session storage until the browser closes. Voice input uses Chrome's Web Speech API, which sends audio to Google.
The catch: it also solves captchas, ticking checkboxes and answering image grids with a vision session, which acts against another site's security control and may break that site's terms. Windows and Android support are experimental, six of the nine agent CLIs have not completed a full side-panel conversation, Firefox lacks the nine tools built on Chrome's debugger, and the install paths for the desktop apps are a `curl | sh` or `irm | iex` one-liner. npm counted 221 downloads in the week to 19 November 2026.
Our take: treat Browsentic like handing a capable stranger your logged-in laptop while you watch; it suits a developer who keeps the approval gate on and uses a separate browser profile, and anyone wanting unattended automation will be better served by a sandboxed browser such as browser-use in its own container.
Where it earns its place
A freelance developer on a Claude Pro or Max plan
Have an agent pull invoices from three logged-in supplier portals each month
Install with `npx browsentic@latest setup`, add the extension, and give the side panel the task across the three tabs; each download stops for approval. No extra cost beyond the existing Claude subscription.
A front-end engineer using Cursor
Let the editor's agent check a change in a staging site behind single sign-on
Register `browsentic mcp` as a stdio MCP server; the agent navigates and reads console output in the already signed-in browser. Form submissions are denied to MCP clients unless you waive the gate in config.
An operations analyst at a small agency
Fill in the same web form for each row of a spreadsheet
Show the agent the task once, keep the approval gate on, and confirm each submission from the side panel card; a dedicated browser profile keeps personal sessions out of reach.
Before you commit
- Setup
- `npx browsentic@latest setup` on any platform with Node.js 20+, or the macOS and Windows one-line installers; setup opens the extension's store page, and the extension pairs with the bridge by itself or with a code.
- Integrations
- Side panel runs Claude Code, Codex or Antigravity, plus Mistral Vibe, Grok Build, Cursor CLI, Qwen Code, OpenCode and DeepSeek Harness in beta; MCP server over stdio for any MCP client; experimental Chrome on Android over adb.
- Data and privacy
- No account, cloud component or telemetry, per the README. Bridge binds to 127.0.0.1 on port 8765, 8766 or 8767. Held secrets live unencrypted in extension session storage until the browser closes; approval grants are kept in `~/.browsentic/approvals.json`.
- Cost at scale
- No paid tier. Cost is your agent CLI's subscription or usage; the limits page allows three concurrent runs by default across eight tab sessions, with a ceiling of eight.
- Limits
- Needs a real, open browser; no headless mode. Recordings stop at 15 minutes. Firefox lacks nine debugger-based tools. Prompt injection is mitigated, not prevented.
What’s good
- Uses the agent CLI subscription you already have; no API key or account
- Approval gates on submissions, uploads, downloads and off-site navigation
- Limits and security pages state the trust boundary and gaps plainly
- Same bridge works from the side panel and as an MCP server
What’s not
- Works in your real signed-in profile; any local process running as you can drive it
- Captcha solving may breach sites' terms of service
- Single maintainer; Windows, Android and six agent CLIs still experimental or beta
- Credential sealing is a filter, and voice input streams audio to Google
Best for
Developers who already pay for Claude Code or Codex and want an agent to work inside sites they are logged in to, with a person approving each consequential step.
Browsentic alternatives
Browsentic: common questions
- Is Browsentic worth it?
- Browsentic is the most security-conscious way we have read about to let an agent loose in your logged-in browser, and it is still a one-person project holding the keys to every session you have. We rate it 3.7 out of 5.
- How much does Browsentic cost?
- Free and open source (Apache-2.0). You pay for whichever agent CLI subscription runs the side panel.
- What are the downsides of Browsentic?
- Works in your real signed-in profile; any local process running as you can drive it; captcha solving may breach sites' terms of service; single maintainer; Windows, Android and six agent CLIs still experimental or beta; credential sealing is a filter, and voice input streams audio to Google.
- Who is Browsentic best for?
- Developers who already pay for Claude Code or Codex and want an agent to work inside sites they are logged in to, with a person approving each consequential step.
- How much does Browsentic cost for a team or in production?
- No paid tier. Cost is your agent CLI's subscription or usage; the limits page allows three concurrent runs by default across eight tab sessions, with a ceiling of eight.
- Is Browsentic safe for company data?
- No account, cloud component or telemetry, per the README. Bridge binds to 127.0.0.1 on port 8765, 8766 or 8767. Held secrets live unencrypted in extension session storage until the browser closes; approval grants are kept in `~/.browsentic/approvals.json`.
- What are the best Browsentic alternatives?
- The closest alternatives we have reviewed are WebBrain, Browser Use, PinchTab, agent-browser and Magentic-UI (MagenticLite).
Answers checked Nov 20, 2026.