NTXM Logo
NTXM BROWSER MINI · v0.1.3 · 127.0.0.1:9090

NTXM Browser.The browser an agent attaches to.

NTXM Browser Mini is a desktop Chromium (CEF) that stays open on your machine — cookies, logins, and tabs included. An agent does not eat HTML or screenshots by default. It gets a small action map, acts by a short id, and gets the next map back in one call.

Version 0.1.3 · Windows & macOSPOST /v1/observe · /v1/act · /v1/extractMCP /mcp · experimental

Unsigned Binary NoticeAs a solo developer, my installers are currently unsigned. Windows Defender, SmartScreen, or your browser may show a "Windows protected your PC" warning or falsely flag the file as harmful. To proceed, click "More info" → "Run anyway". Use at your own risk.

NTXM Browser Mini desktop window

NTXM Browser Mini at a glance

A desktop browser built on Chromium that AI agents and scripts can drive over plain HTTP. An agent asks for a short map of the page, acts on elements by id and gets data back as JSON.

Who it is for
Developers building AI agents that need to use real websites in a browser window they can watch.
Platforms
Windows 10 and 11 (64-bit) and macOS.
Price
Free download.

What it does

  • HTTP API at /v1 on your own computer: observe, act, extract
  • Small page map instead of raw HTML or screenshots
  • Results returned as JSON
  • Your browser session stays on your machine
  • Experimental MCP endpoint

Guides and comparisons

Made by NTXM, built by Nitiksh Kumar.

NTXM Browser is not another Chrome and not a cloud browser farm. It is the local agent browser: one warm profile, HTTP on port 9090, three tools.

What it does

Attach over HTTP /v1

An agent — script, NTXM Agent later, anything that can POST JSON — calls /v1. Sessions are s_19, not window=1&tab=4. One session is one tab. Delete closes that tab only.

01

You run NTXM Browser Mini

Normal desktop window. Your Google / airline / site logins live in that profile.

02

Observe

Ranked controls, short ids (a3f), a few facts, hard cap 800 tokens.

03

Act

fill / click / select / … in one request, optional wait — always returns a fresh observe.

04

Extract

You send a JSON schema; the browser fills it from the page. No LLM inside NTXM Browser.

How it works

A browser with a tiny local API

NTXM Browser Mini is a Chromium browser that stays open on your computer and takes orders over plain HTTP. When it runs, it listens on 127.0.0.1:9090. Your agent or script sends JSON to three endpoints.

POST /v1/observe

Returns a short ranked list of the controls on the page, each with a stable id like a3f, plus a few page facts. Capped at about 800 tokens. No raw HTML or screenshots unless you ask.

POST /v1/act

Click, fill or select by id. It can wait for the page to settle, then sends back a fresh observe so the agent sees the result right away.

POST /v1/extract

Send a JSON schema, get the page's tables or fields back as JSON. No model runs inside the browser.

  • Real desktop profile: cookies, logins and tabs stay between calls.
  • Each session is a tab, so one agent can keep its own tab open.
  • Optional allowlist keeps navigation on the sites you choose.
  • MCP endpoint at /mcp is there too, but still experimental.
Why this exists

One job. That's the product.

Playwright, Playwright MCP, Stagehand, and Browser Use already drive browsers. NTXM Browser does not try to beat them at fleets, stealth, or “click the blue button” on a server.

Attach any model to the Chrome that is already open, using a tiny map and your existing cookies.

That is the whole product.

0.1.3 — what shipped

Production door is /v1

Legacy /windows, /tabs, /page_map, /click still answer for old scripts. They are internal. Manifest and docs lead with /v1.

POST /v1/session*
Create / go / read / delete a tab; host allowlist
POST /v1/observe
Token-capped map + facts + rev; optional diff via since
POST /v1/act
Verb chain + wait; returns observe; typed errors
POST /v1/extract
Schema → JSON (tables/forms only; else extract_failed)
GET /v1/manifest
Machine description of the three tools
GET|POST /mcp
Experimental JSON-RPC: initialize / tools/list / tools/call
Async act wait (tab keeps painting)Short stable ids + reborn_id on remountblocked off-allowlistact navigate via CEF LoadURL after allowlistInstaller NTXM_Browser_Mini_Setup_0.1.3.exe
Measured · honest

Kitchen-sink fixture only.

Measured 2026-09-25 on the shipped 06-kitchen-sink.html only, base http://127.0.0.1:9090. Google Flights and live login traces are not part of this release pack.

7 ms
Warm observe

Target < 300 ms

9 ms
Act (click/fill) → observe

Target < 1.5 s

142
Observe tokens

Target ≤ 800

Deliberately not built

So you stop asking for the wrong product

No server-side natural language (act("click the blue button")) — the caller’s model picks the id

No cloud browsers, proxies, or CAPTCHA farm

No screenshot/HTML as the default agent path

No Playwright / CDP requirement

No tab groups

MCP is not a full SDK and not required to use the product

How an agent talks to it

Production is /v1

MCP (experimental): same three tools, JSON-RPC envelope at /mcp. Use it only if the host cannot call /v1. Cursor-style attach is not the 0.1.3 guarantee.

Allowlist: first successful origin locks the session. Navigate outside → blocked. A tunnel does not widen that.

Production endpointsLocal

POST http://127.0.0.1:9090/v1/session

POST http://127.0.0.1:9090/v1/session/s_19/go

POST http://127.0.0.1:9090/v1/observe

POST http://127.0.0.1:9090/v1/act

POST http://127.0.0.1:9090/v1/extract

GET http://127.0.0.1:9090/v1/manifest

vs Puppeteer · Selenium · Playwright

A running browser, not a library.

Puppeteer, Selenium and Playwright are libraries that launch and drive a browser from your code. NTXM Browser Mini is a browser that is already open, and your agent talks to it over HTTP. The two overlap, but they solve different problems.

NTXM Browser MiniPuppeteer / Selenium / Playwright
How you control itPlain HTTP on 127.0.0.1:9090. Any language that can send JSON.A library in your language, talking to the browser over CDP or WebDriver.
Browser stateNormal window with your real profile. Logins and tabs stay.Often a new context or headless browser. Logins need saving and restoring.
What the agent seesRanked controls with short ids, about 800 tokens max.Full DOM, accessibility tree, screenshots, CSS or XPath selectors.
SetupInstall one app and start it.Install the library, browser binaries and drivers, and manage versions.
Best forAI agents that need a warm, signed-in browser at low token cost.Testing, scraping and general automation.

Why it helps

Built for AI agents

The agent says “click a3f” instead of guessing a CSS selector or reading a huge page dump. Fewer tokens, fewer misses.

Any language

No Python, Node or Java binding needed. If it can make an HTTP request, it can drive the browser.

Session that stays

Same window, same profile. The agent can go straight to your dashboard or a booking page without logging in again.

Less noise

Observe, act, observe. Each step is small and predictable, so the model has less to filter.

Where the classic tools are better

  • Puppeteer, Selenium and Playwright are mature and better for serious testing and large scraping jobs.
  • They give you screenshots, full DOM access and network interception. Browser Mini holds most of that back on purpose to keep the interface small.
  • Browser Mini is early (v0.1.3) and minimal by design. It is not a testing framework.

In short: use NTXM Browser Mini when an AI agent needs a real, signed-in browser with small, stable ids and almost no setup. Use Playwright, Selenium or Puppeteer when you need a full testing or automation library.

Compare

Short. Accurate.

NTXM Browser is the hands. It is not the planner. A future NTXM Agent can call /v1; that is a different product.

NTXM Browser MiniPlaywright / MCPStagehandCloud agent browsers
What it isLocal desktop Chromium + /v1Library / MCP toolsNL act/extract on PlaywrightRented Chrome + agent
SessionYour profile, already logged inUsually a fresh contextYour script’s browserTheir machine
Default see~hundreds of tokens, short idsAccessibility snapshotModel + observeVaries; often heavier
Default extractSchema, no NTXM LLMYou parseTheir model + schemaTheir agent
You installOne installer (Windows or macOS), port 9090Node + browsersSDK + keysAccount + hours
Use cases

What people use it for

Agents on signed-in sites

Let an agent use flight search, dashboards or booking pages with the session you already have.

Cheaper agent loops

Small page maps instead of screenshots or full page trees, so each step costs fewer tokens.

Scripts in any language

Drive it from curl, Python, Node, Rust or PowerShell. No SDK or driver to install.

QA on a warm browser

Run checks against a browser that stays open, instead of starting a new headless context each time.

Stay local

Everything runs on your own computer, not in a rented cloud browser that loses your logins.

Who it is for

  • You want an agent to use your logged-in desktop browser
  • You want HTTP, not a driver stack
  • You care about tokens per step more than pixel computer-use

Not for

  • Replacing Playwright tests
  • 1,000 stealth sessions
  • “Just do the task” with no API

Requirements

  • Windows 10/11 (64-bit) or macOS
  • Local port 9090
  • Windows may show a SmartScreen prompt on first run

Privacy: local-first, no NTXM telemetry. The page the agent opens is between you and that site.

Questions Answered

It lets an AI agent or a script use real websites in a desktop browser on your computer. The agent sends plain HTTP requests to /v1 to see the page, click or type, and get data back as JSON.
Yes. It is a free download for Windows 10/11 (64-bit) and macOS.
Those are libraries that launch and drive a browser from your code, often a fresh one. NTXM Browser Mini is a browser that is already running with your profile, and you drive it with plain HTTP calls. The agent gets a small map of short ids instead of the full DOM or screenshots.
No. Anything that can send a JSON request works: curl, Python, Node, Rust, PowerShell. Start the app and call http://127.0.0.1:9090/v1.
Yes. It uses a real desktop profile, so cookies, logins and open tabs stay there between calls. The agent can go straight to a signed-in page without logging in again.
Official: NTXM Browser Mini. In prose: NTXM Browser. Same app.
Start NTXM Browser Mini. Call /v1. Do not start with /tabs or raw HTML — that is leftover internal REST.
No. /content exists internally. Agents should observe → act → extract.
Playwright is the standard driver. NTXM Browser is a running desktop session with a small agent contract. Different job.
/mcp implements initialize / list / call as a subset. Treat it as experimental. Production integration is /v1.
The engine can open real sites. This release only publishes kitchen-sink benches and traces. Flights/login are not claimed in 0.1.3 notes.
Out of date. 0.1.3 is CEF + /v1 agent contract. Optional bearer auth may exist in settings; the public story is allowlist + local bind.

Local agent browser. /v1 is the door.

NTXM Browser Mini 0.1.3 — TechXcelerate · NTXM

v0.1.3 · Windows .exe · macOS .dmg · See all releases