browsentic Files

Browsentic now runs on seven agent CLIs, reads every file you attach in a short-lived session of its own before your agent sees it, and deals with a captcha wherever it sits on the page. 0.7.1 to 0.7.3 were never published on their own, so this release carries all four steps, newest first.

The extension and the daemon speak a new protocol (18), so update both together: npx browsentic update, then ↻ on the Browsentic card. The Mac app does both from its Overview tab.

0.7.4 · Captchas, in any frame

• page_getPageInfo says when a captcha is on the page. A new captcha field names the vendor, whether it is solved and what to do next, read from every frame without the debugger. The agent no longer waits to be told the word "captcha" before it acts (#30)
• The widget is found at any depth. Turnstile and reCAPTCHA are reached through nested cross-site frames and closed shadow roots, scrolled into view, and waited out while Turnstile says Verifying… (#30)
• Image challenges are answered by a short-lived analyst, one per round. reCAPTCHA 3×3 and 4×4 grids, including the ones that swap in a new picture per tile, and hCaptcha's "tap on…" canvas. It runs on an agent that can see images, Claude Code today. When it cannot finish, the challenge comes back to the caller as an image, answered with the new tiles, points or reload inputs (#30)
• One approval covers every round of a run, and timeoutMs now defaults to 120 s, up to 300 s (#30)

0.7.4 · Attached files, read before your agent sees them

• Each file gets a detached analyst. A one-shot session of your active agent reads that one file, returns a structured report, and is stopped the moment it has answered. It saves no session and leaves nothing running (#31)
• What it cannot read is turned away before anything starts. The kind comes from the file's bytes, not its name: text up to 5 MB on every agent, PDFs up to 10 MB and PNG, JPEG, GIF or WebP up to 5 MB on Claude Code. ZIPs (Word, Excel and PowerPoint files included), other binaries, empty files and anything over 10 MB are rejected with a reason (#31)
• The report travels once, inside your next message, to the conversation the file was attached in, instead of into every conversation's system prompt on every turn. A message sent while the file is still being read waits for it (#31)
• The chip shows the summary, a rejection's reason in amber, or a failure in red with Retry. Removing the chip or ending the conversation cancels the analyst (#31)

0.7.3 · Qwen Code, in beta

• Qwen Code (qwen), beta, the seventh agent. A run starts in --safe-mode, so none of your own hooks, extensions, skills, MCP servers or permission rules reach it, and its startup line is read back: a tool or server Browsentic did not ask for stops the run before the model has spoken (#28, closes [#26])
• Qwen's bundled browser-use and computer-use skills are denied, because the first drives your browser through an extension of its own. Configure a model provider first: Qwen OAuth's free tier has ended (#28)

0.7.2 · Cursor CLI, in beta

• Cursor CLI (cursor-agent), beta. A run gets deny rules that win over every allow, your own included, and a read-only sandbox on macOS and Linux. Every MCP server you set up for Cursor itself is denied by name for the run, and the file they live in is only read, never written (#27, closes [#25])
• The containment check now reads what each runner's config file says, not only that it was written, which also covers the files Antigravity, Mistral Vibe and Grok Build are held by (#27)

0.7.1 · Codex and Mistral Vibe find the browser

• Codex reads the page instead of searching the web. Codex hides MCP tools until the model searches for them, and the setting that switched its own web search off had quietly stopped existing, so it answered from a search. A Codex run is now told where the browser tools are and how to reach them, and web search stays off unless it is mapping a site (#24)
• Codex reads a dense page whole. Its tool result cut goes from 10,000 to 25,000 tokens, the ceiling Claude Code has (#24)
• Sub-agents, goal memory, connector apps and plugin suggestions are off in a Codex run. A sub-agent would act outside the run's approvals (#24)
• Mistral Vibe follow-ups reach the browser. Vibe re-reads a resumed session from the folder it began in, so every turn after the first called the browser with a finished run's id and was refused. A conversation now keeps one folder, rewritten each turn. Start a new conversation after updating: one begun before keeps pointing at its first folder (#24)
• A conversation's folder is renewed every turn, so the daily sweep no longer takes it away mid-conversation (#24)

Documentation

• The README opens with the browsers and agents it works with as marks, leads the quick start with npx browsentic setup and leaves the rest to the install guide, and links every capability to its own page
• The README and the guide name seven agents, with a section each on what Cursor CLI and Qwen Code need first. The tools, errors, limits and troubleshooting pages cover captchas, attached files and the Codex and Vibe fixes

Install

macOS — installs the app, and the app installs everything else, Node included:

:::sh
curl -fsSL https://browsentic.com/install.sh | sh

The Browsentic-0.7.4.dmg below is the same app. It is not notarized yet, so a downloaded copy is blocked on first open: System Settings → Privacy & Security → Open Anyway. The line above avoids that.

Any platform — macOS, Windows or Linux, with Node.js 20 or newer:

:::sh
npx browsentic@0.7.4 setup

That installs the extension, starts the daemon and prints a pairing code. Then load the folder it names at chrome://extensions with Developer mode on, and paste the code into the popup.

Firefox — 140 or newer:

:::sh
npx browsentic@0.7.4 setup --browser firefox

That prints the link to browsentic-0.7.4-firefox.xpi below, signed by Mozilla, and a pairing code. Open the link in Firefox, accept both prompts, and paste the code into the popup. Firefox updates the add-on by itself from then on. If the command says the file is not attached yet, Mozilla is still signing it.

Already running an older version? npx browsentic update refreshes the command and the extension in place and keeps your browser paired. Press ↻ on the Browsentic card afterwards; Firefox needs nothing pressed, and the Mac app offers the update on its Overview tab.

Full changelog: https://github.com/imshaikot/browsentic/compare/v0.7.0...v0.7.4