AI NewsDev toolsAnnouncement

Turbo-Browser-MCP batches Playwright calls so AI agents use fewer tokens

Turbo-Browser-MCP is an MIT Playwright MCP server. It groups form fills, DOM snapshots and tab control into ten batched tools.

AI News

Editorial2 min read

LinkedInX
GitHub social card for the Tthanh2k6/turbo-browser-mcp repository

Image: GitHub

Why it mattersA team paying for every agent call can swap a chatty browser tool for one. The new tool fills a form in one turn instead of eleven.

Giving a coding agent a browser costs a round-trip to the model for every click and field. The agent also pays for the growing chat history on each turn. Turbo-Browser-MCP groups those steps into ten batched tools. Trung Thanh Tran, a developer in Ho Chi Minh City, posted it to Show HN.

The repository is at Tthanh2k6/turbo-browser-mcp under the MIT licence. It had 85 stars, 5 forks and 41 Hacker News points at the time of this item. The current release is v0.5.0, with 407 tests in 86 suites. Node.js 20 or later is required.

What the ten tools do

The author lists them in the README. tabs_batch opens parallel tabs with isolated contexts. inspect_compact returns delta DOM snapshots with a token budget. smart_fill_form fills a form in one call and uploads files from disk. act_fast runs 16 interactive actions such as click, type and press. batch_actions chains multi-step flows. extract_clean_text returns Markdown or plain text. evaluate_js runs scoped JavaScript. page_diagnostics reports console and network state. capture_page takes screenshots and PDFs. network_session blocks resources. A wait_for_human hand-off is also exposed. The README says the server rolls up 25 to 33 Playwright primitives.

The numbers, each one the author's own

Trung Thanh Tran ran six tests with the o200k_base tokeniser and reports them as his own figures. A dense UI snapshot of a 140-button console fell from 18,654 tokens to 1,135 tokens, a 93.9 percent cut. A post-click delta fell from 3,582 tokens to 56 tokens. A 10-field form took 11 turns and 72,614 cumulative tokens on the baseline. The same form finished in one turn and 811 tokens on Turbo-Browser-MCP. A 48 KiB PDF upload cost 44,728 tokens as Base64 and 85 tokens as a disk path. A six-step UI flow went from 3,162 tokens across six turns to 406 tokens in one turn. Opening six tabs in parallel took 420 milliseconds, against 998 milliseconds one at a time.

The guards the author ships with it

The Submit guard is on by default. A form is prepared but not sent until a human unlocks it with allowSubmit=true. File reads and writes are pinned to a workspace path by the TURBO_BROWSER_SAVE_ROOT setting. A submission ledger with idempotency keys stops the same form posting twice. A network guard list and modal-button detection are both on by default. A Hacker News commenter asked the author for a benchmark against browser-use and anomalyco's browse. Those comparisons are not in the README.

Teams running their own agents on the API price their work in tokens per task. A tool that saves most of those tokens on a form fill is worth a repeat on your own site. The guard defaults are the second thing to notice. A browser tool that submits a form before a human looks at the page can run up the wrong bill.

Source

The repository is at Tthanh2k6/turbo-browser-mcp under the MIT licence. The Show HN discussion is on Hacker News.

SourceGitHub

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project