by Trung Thanh Tran
Batched Playwright browser control for agents, with a default submit guard
Pick it if your coding agent runs long browser tasks or fills forms and you want fewer tool calls plus a default stop before the final click. It is the wrong choice for hosted browsers, stealth or Lighthouse audits. Today it needs Node 20 and a git clone, and the project is only days old.
One call can run a whole multi-step flow and return only the elements that changed at the end.
Main trade-off
A brand-new project with one maintainer, no npm package yet and benchmarks that nobody else has run.
MIT licence. Free to clone and run on your own machine. The repository lists no paid tier and no hosted service.
git clone https://github.com/Tthanh2k6/turbo-browser-mcp.git && cd turbo-browser-mcp && npm install
Exposed as MCP tools over stdio (tabs_batch, inspect_compact, smart_fill_form, act_fast and more) that the agent calls by name once the server is registered in its MCP client config.
The permissions this connector asks for, as its own documentation states them.
11 actions an agent can call. Open one for its parameters and an example.
Example call
{"name":"tabs_batch","arguments":{"operations":[{"action":"open","url":"https://example.com/app","isolatedContext":"acct-a"},{"action":"open_many","urls":["https://example.com/docs","https://example.com/pricing"],"concurrency":4}]}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| operations | array | Optional | Operations executed in order. Each has an action plus fields such as url, urls, tabId, isolatedContext, width, height, colorScheme, userAgent or waitUntil. |
| operations[].concurrency | numberdefault 5 | Optional | Tabs opened at once by open_many, from 1 to 10. |
| operations[].isolatedContext | string | Optional | Name of a separate context with its own cookies and storage, for the open action. |
Returns
| Field | Type | Description |
|---|---|---|
| operations_result | object | Outcome of each operation in order, with the tab ids that now exist. |
Example call
{"name":"inspect_compact","arguments":{"mode":"interactive","onlyChanged":true,"maxChars":3000}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| onlyChanged | booleandefault false | Optional | Leave out elements unchanged since the previous snapshot. |
| mode | stringdefault interactive | Optional | One of interactive, forms, article or all. |
| query | string | Optional | Keep only elements matching this regex or substring. |
| maxChars | numberdefault 4000 | Optional | Cap on the snapshot length. |
| includeFrames | booleandefault true | Optional | Look inside child iframes. |
| boxes | booleandefault false | Optional | Add bounding boxes to each snapshot line. |
| savePath | string | Optional | File path inside the working directory or save root where the full snapshot is written. |
Returns
| Field | Type | Description |
|---|---|---|
| snapshot | string | A header with tabId, character count, truncation flag and milliseconds, then one line per element with a short [eN] ref. |
Example call
{"name":"smart_fill_form","arguments":{"fields":[{"selector":"#full_name","value":"Alex Nguyen"},{"selector":"#country","option":"United States"},{"selector":"#agree_terms","check":true},{"selector":"#resume_pdf","file":"./assets/resume.pdf"}],"stopBeforeSubmit":true,"returnSnapshot":true}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| fields | array | Optional | Fields to fill. Each targets a ref or selector and sets value, option, values, check, keys, file or files; frameIdx and all reach into iframes. |
| gmail | object | Optional | Fills a Gmail compose dialog with to, subject and body. |
| overwrite | booleandefault false | Optional | Replace values already present in a field. |
| stopBeforeSubmit | booleandefault true | Optional | Stop before any submit action. |
| returnSnapshot | booleandefault false | Optional | Add a form snapshot to the result. |
| snapshotMaxChars | numberdefault 2500 | Optional | Cap on the returned snapshot. |
Returns
| Field | Type | Description |
|---|---|---|
| fill_report | object | Per-field outcome with the tabId, and a form snapshot when returnSnapshot is true. |
Example call
{"name":"act_fast","arguments":{"action":"fill","selector":"#email","value":"alex@example.com","returnSnapshot":true}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| action | string | Required | One of the 16 actions listed in the description. |
| ref | string | Optional | Compact element ref from a snapshot. selector, text, label or x and y can target instead. |
| value | string | Optional | Text or option to enter for fill, type or select. |
| modifiers | array | Optional | Keys held during the action: Alt, Control, ControlOrMeta, Meta or Shift. |
| allowSubmit | booleandefault false | Optional | Permit an action that looks like Submit, Publish, Send or payment. |
| returnSnapshot | booleandefault false | Optional | Attach a compact snapshot to the response. |
| timeoutMs | number | Optional | Timeout for this action in milliseconds. |
Returns
| Field | Type | Description |
|---|---|---|
| action_outcome | object | The action outcome. A blocked submit reports blockedBy as label, modal, network or ledger. |
Example call
{"name":"batch_actions","arguments":{"steps":[{"action":"smart_fill_form","fields":[{"value":"Alex","selector":"#name"}]},{"text":"Next","action":"click"},{"action":"press","key":"Enter"}],"allowSubmit":true,"expect":{"textMatches":"Receipt\\s*R-\\d+","timeoutMs":8000},"submissionKey":"portal:2026-11-15:entry-1"}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| steps | array | Required | Ordered steps: page actions or nested smart_fill_form steps. Must not be empty. |
| continueOnError | booleandefault false | Optional | Keep going when a step fails. |
| allowSubmit | booleandefault false | Optional | Permit steps that look like a final submit. |
| expect | object | Optional | Proof to look for after an allowed submit: urlMatches, textMatches, selector and timeoutMs. |
| submissionKey | string | Optional | Idempotency key written to a ledger; a repeat with the same key is refused. |
| allowDuplicate | booleandefault false | Optional | Let a repeated submissionKey through. |
| returnSnapshot | booleandefault false | Optional | Attach a snapshot taken after the last step. |
Returns
| Field | Type | Description |
|---|---|---|
| step_outcomes | object | Outcome of every step. After an allowed submit it adds submitVerified and evidence: a new URL, matched text or a selector. |
Example call
{"name":"extract_clean_text","arguments":{"format":"markdown","maxChars":6000}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| format | stringdefault text | Optional | text or markdown. |
| selector | string | Optional | Limit extraction to one element. |
| includeFrames | booleandefault true | Optional | Include text from iframes. |
| maxChars | numberdefault 6000 | Optional | Cap on the returned text. |
| savePath | string | Optional | Write the full text to this file, which must sit inside the save root. |
Returns
| Field | Type | Description |
|---|---|---|
| text | string | A header with tabId, character count and truncation flag, then the extracted text. |
Example call
{"name":"evaluate_js","arguments":{"expression":"(el) => el.textContent","selector":"h1"}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| expression | string | Required | The JavaScript expression or function to run. |
| frameIdx | numberdefault 0 | Optional | Index of the frame to run in. |
| selector | string | Optional | Element to pass as el. ref or text can target instead. |
| arg | object | Optional | Any JSON value passed to the function as arg. |
| savePath | string | Optional | Where to store the result; the path is confined to the save root. |
Returns
| Field | Type | Description |
|---|---|---|
| value | object | The evaluated value as JSON with the tabId, cut at about 20,000 characters. |
Example call
{"name":"page_diagnostics","arguments":{"level":"errors","statusFilter":"errors"}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| level | stringdefault errors | Optional | errors or all console messages. |
| statusFilter | stringdefault errors | Optional | errors or all network requests. |
| includeStatic | booleandefault false | Optional | Include 2xx images, fonts, stylesheets and media when statusFilter is all. |
| urlFilter | string | Optional | Keep requests whose URL matches this substring or regex. |
| reqId | string | Optional | Request id to inspect for headers, post data and response body. |
| part | string | Optional | With reqId, return only request-headers, request-body, response-headers or response-body. |
| dialogPolicy | object | Optional | How to answer page dialogs: action dismiss or accept, with optional promptText. |
| clear | booleandefault false | Optional | Clear the buffers after reading. |
Returns
| Field | Type | Description |
|---|---|---|
| diagnostics | object | Console entries, network requests and failures, downloads and dialog events. |
Example call
{"name":"capture_page","arguments":{"fullPage":true,"savePath":"./shots/home.png","mode":"screenshot"}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| fullPage | booleandefault false | Optional | Capture the whole scrollable page. |
| mode | stringdefault screenshot | Optional | screenshot or pdf. |
| type | stringdefault png | Optional | png or jpeg. |
| quality | numberdefault 80 | Optional | JPEG quality. |
| scale | stringdefault css | Optional | css or device pixel scale. |
| selector | string | Optional | Capture one element. ref or text can target instead. |
| savePath | string | Optional | Where the image or PDF is written, within the save root. |
| returnImage | boolean | Optional | Return the image inline as well. |
Returns
| Field | Type | Description |
|---|---|---|
| file | object | Metadata for the saved file, plus the image itself when returnImage is set. |
Example call
{"name":"network_session","arguments":{"preset":"turbo"}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| preset | string | Optional | turbo, no-images, no-media or clean (clean unblocks everything). |
| blockResources | array | Optional | Resource types to block. |
| blockUrls | array | Optional | URL patterns to block. |
| cookieAction | string | Optional | get, set or clear cookies. |
| storageAction | string | Optional | get, set or clear web storage. |
| storageType | stringdefault local | Optional | local, session or both. |
| saveStorageStatePath | string | Optional | Save storageState JSON to a path inside the save root. |
| loadStorageStatePath | string | Optional | Load storageState JSON from a path inside the save root. |
Returns
| Field | Type | Description |
|---|---|---|
| rules_and_state | object | Active blocking rules and any cookies or storage that were read. |
Example call
{"name":"wait_for_human","arguments":{"message":"Please sign in and finish the security check, then leave the tab open.","selector":"#account-menu","timeoutMs":120000}}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| message | string | Optional | Instruction shown for the person. |
| urlMatches | string | Optional | Resume when the page URL matches this pattern. |
| selector | string | Optional | Resume when this element appears. |
| textMatches | string | Optional | Resume when the page text matches. |
| textGone | string | Optional | Resume when this text disappears. |
| timeoutMs | numberdefault 120000 | Optional | How long to wait; keep it below your client's request timeout. |
Returns
| Field | Type | Description |
|---|---|---|
| wait_result | object | Whether the condition was met before the timeout, with the tabId. |
Turbo Browser MCP is a no-cost MCP server, under the MIT licence, that gives AI agents control of Chromium through Playwright. It groups browser actions into 11 batched tools, returns compact snapshots with only the changed elements, fills a whole form in one call, and blocks Submit, Publish and Send clicks until the agent explicitly allows them.
Delta snapshots
With onlyChanged set, a snapshot lists just the elements added or altered since the last one, so a click costs tokens only for what moved.
Whole-form fill in one call
One request sets text, selects, checkboxes, rich-text editors, shadow DOM and iframe fields, and a Gmail compose box.
Stop before the final click
The last step of a form is refused until the agent opts in, covering Submit, Publish, Send, Apply and payment buttons.
Proof and duplicate protection
From version 0.5.0 an expect check reports submitVerified with evidence, and a submission key keeps a rerun from filing the same entry twice.
Uploads from disk
A file goes in by its path, so a PDF never passes through the conversation as base64.
Parallel isolated tabs
tabs_batch opens up to ten URLs at once, each optionally in its own context with separate cookies and storage.
Four connection modes
auto, cdp, persistent and launch let an agent reuse a signed-in Chrome, a saved profile or a clean headless browser.
Human hand-off
wait_for_human pauses for a login or two-factor step and resumes in the same tab when a condition is met.
Fill a long web form and stop before the final click
An agent fills every field and uploads a PDF in one call, then waits for a person to approve the Submit button.
Run several logged-in sessions side by side
Open tabs in separate isolated contexts so each account keeps its own cookies and storage during one automation run.
Collect article text with less page weight
Block images and media with a preset, then pull clean markdown from the main content instead of raw HTML.
Triage a broken page
Read only the failed requests and console errors, then open the headers or body of one request to find the cause.
Hand a login or 2FA step to a person
The agent pauses, the user signs in or passes the check in the visible tab, and the run resumes in the same session.
| Skill | Works with | Connects by | Checked |
|---|---|---|---|
| Turbo Browser MCP This page | Claude · Codex · Cursor | MCP | 16 Nov |
| Claude · Codex · Cursor | MCP | 11 Nov | |
| Claude · Cursor · Windsurf · VS Code | MCP | 13 Nov | |
| Claude · Cursor · Windsurf | MCP | 11 Oct |
Nothing. The code is MIT licensed, runs on your own machine and needs no account or API key, and the repository lists no paid plan or hosted service (read 16 November 2026). Your only costs are the tokens your agent's model uses and the computer the browser runs on.
Yes. The repository carries an MIT licence file, and package.json names the same licence. No request quota or rate limit is stated anywhere in the docs, so the practical limits are your machine and the sites you visit.
Not by default. Final actions that look like Submit, Publish, Send, Apply, Confirm or a payment are blocked unless the agent passes allowSubmit or you start the server with --allow-submit. The README admits button labels can mislead, so it adds an optional network guard for POST and PUT requests, switched on with TURBO_BROWSER_NETWORK_GUARD=true and off by default because many multi-step wizards send a request on every Next.
The README shows an npx line, but the npm registry returned package not found when Hokai checked on 16 November 2026, even though package.json and server.json both name the package. Until the maintainer publishes it, clone the repository, run npm install and point your MCP client at src/index.mjs with node. Hokai did exactly that at the newest commit on the same day: the doctor check reported Node, Playwright and a headless Chromium, and a stdio tool-list call returned every tool described on this page.
Yes. In its default auto mode the server first looks for Chrome's remote debugging endpoint at http://127.0.0.1:9222, attaches to that browser, and only then falls back to launching Chromium. A dedicated profile through --launch-chrome avoids the Allow remote debugging prompt that newer Chrome versions show on connection. A site that re-checks for automation on every request will still treat the browser as automated.
Turbo Browser MCP is an open-source Model Context Protocol server by developer Trung Thanh Tran, first published in November 2026, that lets an AI agent drive a real Chromium browser through Playwright. It is a skill, not a tool: you never open it yourself; it is called by Claude Code, Cursor or another MCP client. Its aim is fewer tokens and fewer round trips. Instead of one call per click, it groups form fills, clicks and tab opens into single calls, and after a click it can return only the page elements that changed. The repository showed 106 stars on 16 November 2026.
The server speaks MCP over stdio, needs no account or API key, and runs on Node 20 or newer through playwright-core. It can attach to a Chrome you already have open with remote debugging, reuse a saved profile folder, or start a clean headless browser. Pages come back as compact numbered snapshots, and the server looks inside iframes and open shadow DOM. Files are uploaded from a path on disk rather than pasted into the conversation as base64.
By default the server refuses the last step of a form: buttons that look like Submit, Publish, Send or a payment, and Enter presses inside forms, are blocked until the agent sets allowSubmit, or the server is launched with a startup flag. Version 0.5.0 added a proof check, a duplicate-submission ledger and a modal rule; the maintainer README says readers running agents on real portals reported gaps in label-only checks (read 16 November 2026). The README gives setup for Claude Code, Cursor, Antigravity and Codex; Hokai has not tried other MCP clients.
Hosted browsers with proxies and CAPTCHA solving are what Kernel, Notte and Browserbase sell; this server runs only on your own machine and does not try to hide automation. Mantis is the local rival with the same token-saving aim and answers in numbered text. The maintainer comparison table sets it against Playwright MCP and Chrome DevTools MCP, neither of which has a Hokai page yet; the second keeps Lighthouse audits and heap snapshots that this server leaves out. See the wider browser skills shelf and the best MCP servers by job guide.
Decision: try it on your own machine if you want a local, token-lean browser with a guarded submit and can live with a git clone. Choose a hosted browser if you need proxies, and keep Chrome DevTools MCP beside it if you need Lighthouse.
Protocol, authentication, permissions and the call contracts on this page are read from Trung Thanh Tran's own documentation. The verdict is Hokai's reading of them, not the maintainer's claim.