turbo-browser-mcp: a Playwright MCP server that fills forms and batches browser steps in one call
turbo-browser-mcp is a Playwright MCP server for browser automation that fills a whole form in one call, returns only changed page elements, and blocks final submit buttons by default.
Tthanh2k6/turbo-browser-mcp gives an AI coding assistant a browser it can drive in fewer, larger steps. It is a Playwright MCP server that fills a whole form in one call, returns only the parts of a page that changed, and chains clicks, typing and page reads into a single call, while a submit guard blocks final commit buttons by default.
What is turbo-browser-mcp?
turbo-browser-mcp is a Playwright MCP server for browser automation that is built to cut agent turns and tokens.
Each tool does more per call than the usual one-action-per-call design. smart_fill_form fills a whole form in one call, including React inputs, selects, comboboxes, rich-text editors and file uploads read from disk. inspect_compact can return only the elements that changed since the last snapshot. batch_actions runs a list of clicks, key presses, form steps and page reads as one call. Fewer calls means fewer model turns and less page text replayed, according to the README.
The server has 11 tools in total, covering tabs, compact snapshots, form filling, single and batched actions, clean page text, JavaScript evaluation, diagnostics, screenshots and PDFs, network sessions, and handing a step to a person.
The README measures the saving on local fixture pages in headless Chromium, counting tokens with tiktoken o200k_base, and states that model time is not included and the figures are an upper bound on what an agent gains, not an agent result. On a 10-field form, the README reports 11 turns and 12,197 direct tokens for the step-by-step baseline against 1 turn and 811 tokens with smart_fill_form, a saving of 93.4 percent direct and 98.9 percent cumulative. On a snapshot after a click, onlyChanged reports 168 characters against 12,825 characters for a full ARIA snapshot, a saving of 98.4 percent against ARIA.
Who is turbo-browser-mcp for?
turbo-browser-mcp is for developers who drive a browser through an AI coding assistant and want fewer turns, smaller snapshots and a guard on final submissions.
It works with clients that read an mcpServers JSON config, which the README names as Cursor, Claude Desktop, Antigravity and others, and it has a direct command for Claude Code. It needs Node 20 or newer and a Chromium-based browser, named in the README as Chrome, Edge, Brave, or Playwright's Chromium. By default the server first tries a Chrome already listening on http://127.0.0.1:9222, so the agent can use a browser the reader is signed in to, and otherwise launches its own headless Chromium.
It fits work where an agent reads pages, fills forms, opens several tabs, extracts text, or checks console and network errors. The README also names chrome-devtools-mcp as the tool to keep alongside it for Lighthouse audits, performance traces and heap snapshots, which turbo-browser-mcp does not offer.
Quick facts
- Platform: Node 20 or newer, with a Chromium-based browser (Chrome, Edge, Brave, or Playwright's Chromium).
- License: MIT.
- Language: JavaScript, with the only runtime dependency in package.json being playwright-core at version 1.49.0 or newer.
- Latest release: v0.7.0, published on November 17, 2026, as the newest entry on the Releases page as of November 18, 2026.
- Popularity: about 130 GitHub stars as of November 18, 2026.
- Tools: 11 tools by default; with --tools core only 6 tools are listed and calls to the other tools are refused.
What stands out
- One-call form filling: smart_fill_form fills a whole form in one call, including React inputs, selects, comboboxes, rich-text editors and file uploads read from disk.
- Delta snapshots: inspect_compact can return only the elements that changed since the last snapshot, with a star marking a new ref and a Removed line listing refs that went away.
- Batched steps: batch_actions runs ordered actions, form steps and multi-page navigate and extract steps as one call, and stops when the submit guard blocks a step unless continueOnError is set.
- Submit guard on by default: buttons that look like a final commit, including Submit, Send, Publish, Pay and Confirm and their Korean, Japanese, Chinese and Vietnamese equivalents, are blocked by default, and the guard can be turned off per call with allowSubmit set to true or at startup with --allow-submit.
- Commit checks: batch_actions accepts expect to verify a submission, including textNew for text that was not on the page when the batch started, and submissionKey to refuse a later call with the same key unless allowDuplicate is set.
How do I install turbo-browser-mcp?
turbo-browser-mcp is installed with the npx command for Claude Code or with an mcpServers JSON config for other clients, and the setup can then be checked with the --doctor command.
From the project's README, the commands below are in the README's order. The README is linked in the sources, as is the Releases page.
claude mcp add turbo-browser -- npx -y turbo-browser-mcp{
"mcpServers": {
"turbo-browser": {
"command": "npx",
"args": ["-y", "turbo-browser-mcp"],
"env": { "TURBO_BROWSER_SAVE_ROOT": "/absolute/path/to/workspace" }
}
}
}npx -y turbo-browser-mcp --doctornpx playwright-core install chromiumgit clone https://github.com/Tthanh2k6/turbo-browser-mcp.gitcd turbo-browser-mcpnpm installnode src/index.mjs --doctorTURBO_BROWSER_SAVE_ROOT is the only folder the server may write screenshots, PDFs and storage files to, or read storage files from. Leave it out and the current working directory is used. The package also ships a short Agent Skill, skills/turbo-browser/SKILL.md, that teaches the low-token loop of one inspect_compact call, then onlyChanged, whole forms with one smart_fill_form, and textNew to verify a submit.
Try it with your AI. Try it with your own assistant, using only the official commands and tools named in the README:
Help me try turbo-browser-mcp on my own setup. First show me the risky steps and wait
for me to confirm before you run anything. Use only the official commands and interfaces
named in the README: the Claude Code command 'claude mcp add turbo-browser -- npx -y
turbo-browser-mcp', the mcpServers JSON config with command 'npx' and args ['-y',
'turbo-browser-mcp'], and the check command 'npx -y turbo-browser-mcp --doctor'. Check
that Node 20 or newer is installed and that a Chromium-based browser is available. Then
use only the README tools inspect_compact, smart_fill_form and batch_actions on [the
page I name] to [the task I name, for example fill a form without submitting it]. Keep
the submit guard on by default, do not pass allowSubmit as true, and leave the final
Submit, Send, Publish, Pay or Confirm click to me. Do not install anything from unknown
sources, do not touch credential files, do not delete anything, and do not run a command
I have not confirmed.What are turbo-browser-mcp's limits?
turbo-browser-mcp's limits come from its guards, its short observation windows, and the sites it is used against.
The submit guard checks labels and roles. The README lists three gaps with opt-in answers: a click that landed is not a confirmed submission, so use expect; a re-run is a second entry, not a retry, so use submissionKey; and labels can be wrong, so affirmative buttons inside a dialog count as commits. The network guard is off by default, it is turned on with TURBO_BROWSER_NETWORK_GUARD set to true, and it watches the action plus 250 milliseconds afterwards, so a commit fired later by a timer is not caught. WebSocket messages, service-worker requests and cross-site fetch POST requests are not inspected.
The host allowlist is off by default. When it is set, it limits where the browser may go, but sub-resources load from any host and iframes are not restricted by default, and strict mode is still best effort, so the README says to treat the setting as a guard rail for an agent, not as a security boundary. Secrets are replaced at typing time from a dotenv file and values are never logged or returned, and values shorter than 4 characters are not redacted. Content boundaries are off by default and add about 21 tokens per result. The bridge for Chrome 144 and newer holds a Chrome connection the reader already approved, so every request must carry a token, and --bridge-insecure runs it open, so any program on the machine can then use the approved Chrome session, cookies included.
The server does not hide that the browser is automated: no stealth patches, no fingerprint changes, and no CAPTCHA solving. Sites with bot detection may need wait_for_human and a person to pass the check in the same browser. Portaled comboboxes have a keyboard fallback that searches up to 100 options, and virtualized lists beyond the included fixtures still need testing against real apps. Closed Shadow roots that existed before the server attached to a tab cannot be reached. The agent benchmark has one trial per stack and task so far, so the README states that one trial per cell cannot rank the stacks, and the ten-trial protocol in benchmarks/E2E_PLAN.md has not been run.
For security, the latest minor version on npm gets fixes and older versions do not, and the project is a one-person project with reports handled on a best-effort basis and no promised response time. SECURITY.md asks readers to report a vulnerability privately through GitHub and not to open a public issue for a security problem.
Related on claudefox
- Digital Human Koubo: AI talking-head videos from one photo with an agent skill
- BLA (bla-agents): a CLI and MCP server that trips AI reverse-engineering limits on release files
- How to call Gemini 4 Hydrogen from Claude Code: gemini-api CLI, Docs MCP and skills
- reanimate is a Claude Code skill that clones an animation from a site, video or gif, then measures the copy against the original frame by frame
Problems and feedback
Report problems and request features on the project's Issues page, linked in the sources. CONTRIBUTING.md says the project is a one-person project, that issues and pull requests are welcome, and that replies are best effort.
Source: turbo-browser-mcp repository, README, Releases, v0.7.0 release, LICENSE, SECURITY.md, CONTRIBUTING.md, package.json, Bundled skill, End-to-end benchmark comparison, Issues
This story first appeared in the claudefox daily briefing for November 18, 2026.