jev-ultrafast-mcp

English · 简体中文

Hand the browser work off — an MCP server that drives the page for your agent.

One tool call instead of twenty. Three seconds instead of a minute. A cent instead of a frontier model's context. And it never invents a target: it picks from what the page actually has, and the server refuses rather than guesses.

Quick start

Three commands, then restart your client.

git clone https://github.com/jiawei686/jev-ultrafast-mcp.git
cd jev-ultrafast-mcp
python3 -m venv .venv
.venv/bin/pip install -e .          # Windows: .venv\Scripts\pip install -e .

python scripts/install.py           # finds your MCP clients and writes their config

install.py looks for WorkBuddy, Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code, Cline, Windsurf and Gemini CLI, and writes the format each one expects — merging into your existing config and saving a .bak first. Needs Python ≥ 3.10 and any Chromium-family browser.

Restart the client, and then just say what you want:

You: Set this form to 3 adults, tick Nonstop only, then submit it. Your agent: browser_goal(goal=…, verify=[…]) — one call; the loop runs server-side, and the page is checked by code afterwards. (what that costs)

You: Open example.com and tell me what the page says. Your agent: browser_open → reads the element table → answers. (verbatim run)

Restarted and the tools are not there? Some clients make you approve the server once. In WorkBuddy that is Connectors → Custom connectors → Trust. The approval is remembered against the config itself, so if you later edit the config it asks once more.

Cheap and fast, and here is the bill

A three-step goal on a real page, driven by browser_goal. This is everything your agent sent and everything it got back — one turn, and the page never entered its context:

browser_goal(
  goal="On this flight search form: set Passengers to 3 adults, tick the 'Nonstop only' "
       "checkbox, then submit the search. Do not type into any city field.",
  verify=[{"type": "text_contains", "text": "3 adults · nonstop"}],
)

goal: On this flight search form: set Passengers to 3 adults, …
status: done
steps: 3
turbo: 4 decisions · 14,626 tokens · 1.8s model + 1.1s page · 3.3s wall
trace:
  1. SELECT e6 Passengers → ok (759ms model / 30ms browser)
  2. TOGGLE e7 Nonstop only → ok (336ms model / 692ms browser)
  3. CLICK e8 Search → ok (370ms model / 410ms browser)
  4. DONE (conf 0.93)
verified: PASS
  ok text_contains: '3 adults · nonstop' found in page text

What it cost. A cent for the whole day, and a cent is all of it:

What the second time costs. Nothing. The second time is a recorded macro, and a macro makes no model calls at all — it does not even need a key.

How long it took. 3.3 s wall for the whole goal: 1.8 s of model, 1.1 s of page. Every run prints that line itself, so the numbers are checkable rather than persuasive.

That run is not a mock-up. scripts/turbo_check.py reproduces it against a real Chrome and the real model, and then checks the page with code rather than trusting the model's account of its own work. The three actions — plus the reading and re-reading between them — all happened on the server. Your agent spent one turn and never saw an element table.

The division of labour is the whole design decision, so it is yours to make per task:

python3 -m venv ~/.jev-ultrafast-mcp/venv
~/.jev-ultrafast-mcp/venv/bin/pip install "git+https://github.com/jiawei686/jev-ultrafast-mcp"

python scripts/install.py --list              # what is installed, and the file each one reads
python scripts/install.py --print             # show the config it would write, change nothing
python scripts/install.py -c cursor,codex     # only these two
python scripts/install.py --headed            # keep a visible browser window
python scripts/install.py --allow-domains example.com,*.example.org
python scripts/install.py --uninstall         # take the entry back out

What it is

Browser automation usually makes the agent do the driving: read the page, pick one element, act, read again to see whether that worked. Ten clicks is ten turns, the page passes through the agent's context every time, and a mis-click rarely announces itself.

This server can take that job instead. browser_goal is one tool call from your agent; the loop runs here, server-side, with Jev — TypeSafe's decision model — choosing each step. The model never writes a selector: it picks among the elements the page actually has, and the server refuses anything that is not on the page rather than guessing. When it stops, browser_assert checks the page it left behind in code, and a passing assertion outranks the model's own account of what it did.

Four things follow from that:

• One call, not one per click. The run above took a 3-step goal on a real page through 4 decisions, 14,626 tokens, 1.8 s model + 1.1 s page, 3.3 s wall — for one turn of your agent's context.
• Accurate by construction. A target is a ref from a numbered table of what is on the page, not a selector or a coordinate the model invented, and the action is re-checked against the page before it runs.
• Free after the first run. Record the path once; replay costs zero model calls, works with no key at all, and refuses to proceed when the page no longer matches.
• Text, not pixels. No screenshots, no HTML dumps. It speaks CDP straight to a Chrome you already have — no Playwright, no Selenium, no screenshot pipeline.

Everything except browser_goal — browser_open, browser_observe, browser_act, browser_assert, browser_macro — needs no key, no account, and no network beyond the page itself, from any MCP client: WorkBuddy, Claude Code, Codex, Cursor or VS Code. If you would rather keep your hands on the wheel, that whole surface is still here.

browser_open  →  element table  →  browser_act [refs]  →  browser_assert

Inspired by browser-use/jev-ultrafast and TypeSafe's typed-question API. Independent project, not affiliated with either — see docs/DESIGN.md for what is different and why.

Contents · Quick start · Cheap and fast · What it is · What a session looks like · Connecting an agent · What you can ask it to do · What the agent reads · Why another browser MCP? · Tools · Configuration · FAQ · Try it without an agent · See also

What a session actually looks like

You say:

Open example.com and tell me what the page says.

Your agent does this, and this is everything it sees:

browser_open("https://example.com")
  [obs#1] https://example.com/  "Example Domain"  scroll=0/216  reachable=1/1
  e1   lnk    More information...

browser_observe()
  [delta#2] … 1 element
    = no change (1 element)

Then it answers. No screenshot was taken, no HTML was dumped, and the page never entered a model's context: your agent read the table and answered.

A more realistic one — searching a real site, with your agent doing the driving:

browser_open("https://duckduckgo.com")
  e4   cmb*   Search with DuckDuckGo ▸ ""

browser_act([{type, ref: "e4", text: "python asyncio tutorial"}, {keys, key: "Enter"}])
      → 2/2 ops ok, one round trip, page navigated

browser_observe()
  [delta#3] https://duckduckgo.com/?…&q=python+asyncio+tutorial  reachable=9/59
  + e5   lnk    Python Asyncio Tutorial
  + e6   lnk    Async IO in Python: A Complete Walkthrough
  …
    43 new, 0 changed, 0 gone

browser_assert([{url_contains, text: "q="}, {count_at_least, role: "link", min: 5}])
  PASS

That is a verbatim run against the live web — scripts/live_check.py reproduces it end to end.

Connecting an agent

{
  "mcpServers": {
    "jev-ultrafast-mcp": {
      "command": "/ABS/PATH/jev-ultrafast-mcp/.venv/bin/python",
      "args": ["-m", "jev_ultrafast_mcp"],
      "env": { "JEVMCP_HEADLESS": "1" }
    }
  }
}

claude mcp add --scope user jev-ultrafast-mcp \
  --env JEVMCP_HEADLESS=1 \
  -- /ABS/PATH/jev-ultrafast-mcp/.venv/bin/python -m jev_ultrafast_mcp

[mcp_servers.jev-ultrafast-mcp]
command = "/ABS/PATH/jev-ultrafast-mcp/.venv/bin/python"
args = ["-m", "jev_ultrafast_mcp"]
startup_timeout_sec = 20

[mcp_servers.jev-ultrafast-mcp.env]
JEVMCP_HEADLESS = "1"

{
  "servers": {
    "jev-ultrafast-mcp": {
      "type": "stdio",
      "command": "/ABS/PATH/jev-ultrafast-mcp/.venv/bin/python",
      "args": ["-m", "jev_ultrafast_mcp"],
      "env": { "JEVMCP_HEADLESS": "1" }
    }
  }
}

The browser does not start until the first browser_open, and the tab it drives is a background tab it owns — focus emulation keeps animations and menus running without stealing your window.

What you can ask it to do

What it is not

Being clear about this saves everyone time:

• It never looks at pixels. A captcha, a chart, a canvas-only app — anything that needs real visual judgement — is out of scope. Use a screenshot-and-vision agent for those, or use the screenshot op here to capture evidence for a human.
• It is not a scraper framework. One browser, one session at a time. No proxy rotation, no concurrency, no crawling at scale.
• It is not a recorder for humans. There is no click-to-record UI; macros are recorded by the agent driving the task normally.

What the agent actually reads

Not a DOM dump, not a screenshot — a table of the controls it can act on. Each row is a ref (element number), a role code, flags, and the accessible name; editable things carry their current value, and selectable things carry their options:

[obs#1] http://127.0.0.1:54409/fixture.html  "Ultrafast Fixture"  scroll=0/860  reachable=16/16
e1   lnk    Home
e2   lnk    About
e3   lnk    Open popup
e4   inp*   Where from? ▸ ""
e5   cmb*   Where to? ▸ ""
e6   cmb    Passengers ▸ 1 adult opts{1 adult=1 | 2 adults=2 | 3 adults=3 | 4 adults=4}
e7   chk·   Nonstop only
e8   btn    Search
e10  inp*   Password ▸ ""
e11  file    CV accept=.pdf,.txt
e12  btn    Delete account

Flags: * editable · » off-screen (the server scrolls it into view) · ⊘ covered by something else · ▾ expanded · ✓/· checked state. reachable=16/19 means three controls exist but are covered or off-screen right now.

After an action it reports only what changed — that is the single biggest saving in a long loop:

[delta#2] http://127.0.0.1:54409/fixture.html  "Ultrafast Fixture"  reachable=16/16
~ e4   inp*   Where from? ▸ "Zurich"   (was "")
~ e7   chk✓   Nonstop only
  2 changed, 0 new, 0 gone

An action that accomplished nothing is the most expensive thing in an agent loop, because the model retries it. So it is spelled out in one line:

[delta#3] … 16 elements
  = no change (16 elements)

New rows appear with + and disappear with -. When two controls share a name, the row carries the context that tells them apart:

+ e17  btn    Select  @Zurich → Anywhere Option 1 · 1 adult · nonstop Select
+ e18  btn    Select  @Zurich → Anywhere Option 2 · 1 adult · nonstop Select

And a ref that no longer points at anything is refused, with a reason instead of a wrong click:

[{"op": "click", "ok": false, "ref": "e999", "error": "detached"}]

Why another browser MCP?

Two differences, and the first one is the reason this exists.

The agent is allowed to decline the driving. A browser flow is a loop, and in most servers that loop lives in the calling agent: read the page, name one element, wait, read again. Fine for two steps, absurd for twenty — twenty turns of an expensive context to do what a smaller model could have done in one call. Here you can hand the whole goal over instead and pay a single turn, or keep the wheel and drive it yourself. Same tools, same guards, either way.

The model never invents a target. Most browser MCP servers hand over CDP primitives — click_at_xy, a CSS selector, evaluate. Maximum flexibility, minimum safety: a wrong selector fails silently or, worse, succeeds on the wrong element. Here a target is a ref from a numbered table of what is on the page, turning that ref into a real click is the server's problem, and the server refuses rather than guesses. That is also what makes the handoff safe: whatever is driving is choosing among options the page actually has, so accuracy does not rest on it being careful.

Batching and deltas are not cosmetic. In the bundled end-to-end run, 29 ops and their follow-up observations cost 15.4 KB of context, of which 13.6 KB was deltas and 1.8 KB full tables — the model re-reads only the part of the page that moved.

Tools

Ten tools. Most sessions need four of them.

browser_open(url, session="default", hint="")

Opens a URL in its own tab and returns the full element table. hint restates your goal in one line and is echoed back.

browser_observe(session="default", mode="auto", include_text=True, include_json=False)

Re-reads the page. auto emits a delta; full forces the whole table, delta forces a diff. = no change means the last action did nothing — change strategy, do not retry.

browser_act(ops, session="default", dry_run=False, stop_on_error=True, observe_after=True)

Executes ops in order in one round trip, then returns a delta.

{"ops": [
  {"op": "type",   "ref": "e4", "text": "Zurich"},
  {"op": "select", "ref": "e6", "value": "3 adults"},
  {"op": "toggle", "ref": "e7"},
  {"op": "click",  "ref": "e8"}
]}

A failing op reports why: occluded, detached, target_changed, page_changed, needs_confirmation, blocked_by_policy. Reach for browser_observe, not a retry.

For tabs, prefer target_id over index. Indexes are positional and get renumbered whenever the tab list changes, so an index read one call ago can address a different tab.

browser_assert(checks, session="default")

Deterministic checks — no model judgement about whether it worked.

{"checks": [
  {"type": "url_matches",    "pattern": "*/checkout*"},
  {"type": "text_contains",  "text": "Order confirmed"},
  {"type": "element_exists", "role": "button", "name": "Continue"},
  {"type": "value_equals",   "ref": "e4", "value": "Zurich"},
  {"type": "count_at_least", "role": "link", "min": 3}
]}

browser_macro(action, session="default", name="", params={}, ...)

record_start → drive the task → record_stop → run. Replay costs no model calls: it navigates back to where the task began and re-resolves every step by role + accessible name, refusing weak or ambiguous matches rather than clicking the wrong thing. params fills {{placeholders}} in typed text and URLs.

browser_goal(goal, session="default", max_steps=20, verify=[...])

Runs the whole loop server-side using TypeSafe speculative fan-out (one request per step). Needs TYPESAFE_API_KEY, or OPENROUTER_API_KEY with TYPESAFE_BASE_URL pointed at OpenRouter's decisions route. Returns verified: PASS/FAIL when verify checks are supplied.

Every run also reports its own bill — turbo: 4 decisions · 14,626 tokens · 1.8s model + 1.1s page · 3.3s wall — so what the handoff cost is visible in the answer, alongside how much of the wall time was the model and how much was the page.

Every way the decision model can fail — no key, no credits, unreachable, a malformed answer, a body that is not JSON — comes back as turbo_unavailable: with nothing executed. The trace of the steps already taken is kept, so a run that dies on step five still reports what steps one to four did.

status is the model's own summary, and verify is checked by code, so when the two disagree the assertion decides: if the page passes your checks the run reports status: done whatever the model said, and the trace records that it overruled. This is the ordinary shape of a goal whose last action removes what it acted on — click a check-in button and the button is gone, so the model, finding nothing left to do, reports BLOCKED on a goal that in fact succeeded.

browser_tabs · browser_sessions · browser_close · browser_doctor

Tab management (list / new / switch / close), session listing, teardown, and a self-check that reports which browser was found and whether it is reachable.

Configuration

All optional; the defaults are the point.

JEVMCP_MODE=attach is the "use the browser I already have open" route — the one to take when the login you need already lives in your own profile. Chrome 144+ exposes that through chrome://inspect/#remote-debugging, and its server answers 404 to /json/version by design; jev falls back to DevToolsActivePort rather than treating that as "nothing is listening". Attach mode only ever touches the tab it opens: browser_close detaches rather than quitting, and the same holds when the server exits. Your other windows, and the session in them, are never closed.

Everything above the last seven rows is local: it configures a browser on your machine. Only the decision-model group talks to the network, and only when browser_goal actually runs. Pointing TYPESAFE_BASE_URL at OpenRouter means one OPENROUTER_API_KEY covers both the decision model and the text helper, and needs no TypeSafe account.

Two of these are worth setting before you point an agent at your own accounts: JEVMCP_ALLOW_DOMAINS pins the browser to a set of hosts and refuses everything else, and a persistent JEVMCP_PROFILE_DIR means you log in once by hand instead of teaching the model your password.

FAQ

So this is just another model doing the work? Who is in charge? You are, and you choose per task. browser_goal puts Jev — a small decision model — in charge of one goal: which of the page's elements to touch, one step at a time. It is not a general agent, it has no memory between goals, and it never writes code or selectors, only picks from options the server hands it. Your agent still decides what to ask for, and verify decides whether it actually happened. If you would rather be in the loop for every step, do not call that one tool — nothing else sends anything anywhere.

Do I need an API key or an account? Not for the browser tools. browser_open, browser_observe, browser_act, browser_assert, browser_macro and the tab/session tools never call out — no telemetry, no phone-home, nothing leaves your machine. browser_goal is the exception, and it is opt-in: it sends your goal and the current element table to a decision model, which is why it needs a key. Leave that one tool unused and nothing about the page goes anywhere.

Will a browser window pop up and take over my screen? No. It runs headless by default and drives a background tab it owns — animations and menus still work, but nothing steals focus. --headed (or JEVMCP_HEADLESS=0) shows the window if you want to watch it work.

How do I use it on a site I am logged into? Set JEVMCP_PROFILE_DIR to a persistent directory, open the browser once by hand, log in, and the session is remembered. That is far better than teaching an agent your password — and password fields are redacted in observations when you do type them.

Can it just use the browser I already have open, with my logins in it? Yes. JEVMCP_MODE=attach plus JEVMCP_CDP_URL=http://127.0.0.1:9222 drives your own Chrome. In Chrome 144+ you switch debugging on from chrome://inspect/#remote-debugging — no restart, so your tabs and logins survive — and Chrome asks you to approve the client. The first connection waits on that click, so give it a moment before deciding it failed.

Do I have to approve that click for every action? No. The approval is per browser session, not per connection and not per action. Once you have approved it, every later action rides the same open WebSocket and never prompts again — not even from a fresh process (measured: three new processes connecting twenty minutes after the click, all accepted with no prompt). So the cost is one click per browser session, not one per operation. To drop even that, use the default JEVMCP_MODE=launch: it starts its own browser on a real debugging port with no approval dialog at all, at the price of logging in once in that profile.

curl http://127.0.0.1:9222/json/version returns 404. Is debugging even on? Probably yes. The server behind chrome://inspect/#remote-debugging is WebSocket-only and deliberately serves no HTTP discovery endpoints, so a 404 there is the documented behaviour rather than a broken setup (and it is not the same thing as --remote-debugging-port=9222, even though both print port 9222). jev does not rely on it: when /json/version does not answer, it reads the port and the browser WebSocket path out of Chrome's DevToolsActivePort file. If your browser keeps its data directory somewhere unusual, point JEVMCP_ATTACH_PROFILE_DIR at it.

Nothing is happening and the page looks empty. A page that renders from JavaScript can briefly look empty. The server waits for controls to appear (up to JEVMCP_SETTLE_TIMEOUT), but if a site is stuck behind a cookie wall or a consent dialog, the element table will show it — look for the overlay warning in the observation header.

There is a captcha. Can it solve it? No, and that is deliberate — it never looks at pixels. Use a screenshot-and-vision agent for that.

Is it on PyPI? Is it in the MCP registry? Not yet. pip install "git+https://github.com/jiawei686/jev-ultrafast-mcp" installs exactly what a release would. server.json is already in the repo for the official registry, which publishes a package — so it lands there in the same step as the first PyPI release. See Publishing for the state of that.

How is this different from the Playwright MCP? Playwright's server exposes page primitives; the agent writes selectors and coordinates. This one exposes a numbered table of controls and refuses ambiguous targets. If you need pixel-level control or a mature recorded-testing ecosystem, use Playwright. If you want an agent that cannot silently click the wrong button, use this.

Is this an alternative to browser-use? They solve the same problem from opposite ends. browser-use is a library that runs the agent loop in-process; this is an MCP server that gives your existing agent the same kind of hands. The optional turbo path here is a port of browser-use/jev-ultrafast, which asks a decision model one typed question per step instead of free-form text.

Is it safe to let it loose on my accounts? It is built assuming it should not be trusted. Destructive-sounding clicks come back as needs_confirmation instead of executing, JEVMCP_ALLOW_DOMAINS refuses navigation outside a domain you list, sensitive fields are redacted, and eval is off unless you turn it on. Start with a domain allowlist and an account you do not mind breaking.

Try it without an agent

.venv/bin/python scripts/smoke.py             # headless, 58 checks
.venv/bin/python scripts/smoke.py --headed    # watch it drive

This launches Chrome, serves tests/fixture.html, and drives the real code paths: batch execution, autocomplete, a covering modal, shadow DOM, a same-origin iframe, a file upload, a password field, a destructive-click guard, stale refs, macro record/replay, tab handoff, screenshots, and a page that renders after readyState already says "complete".

1. Observation — one atomic read, indexed refs
  [ok  ] element table is not empty  — 16 elements
  [ok  ] shadow DOM element indexed  — Shadow action -> e14
...
5. Occlusion — precomputed, not discovered by a failed click
  [ok  ] covered control flagged before any click  — e8 occluded=True
  [ok  ] click on a covered control is refused with a reason  — occluded
...
  58/58 checks passed

To test against the real web rather than a fixture:

.venv/bin/python scripts/live_check.py            # Bing + DuckDuckGo + tabs + screenshots
.venv/bin/python scripts/live_check.py --headed   # watch it happen

That one needs the network and third-party sites, so it is deliberately not part of CI. Unreachable sites are reported as skipped, and the summary says so plainly, so a fully-skipped run cannot be mistaken for a passing one.

And to prove turbo mode itself — the one path that spends money, and therefore the one nothing else exercises end to end:

.venv/bin/python scripts/turbo_check.py           # the model drives: dropdown, checkbox, submit
.venv/bin/python scripts/turbo_check.py --headed  # watch it decide

It serves the same fixture, points a real Chrome at it, and lets Jev drive the goal, then verifies the page the model left behind with code rather than trusting its claim of success. Without a key it prints skipped and exits 0, so the exit code and the word agree.

A check-in that stops paying for itself

examples/checkin.html is a stand-in for the thing people actually automate: a daily button. scripts/checkin.py drives it in three stages, cheapest first — and the point is that only the first run ever costs anything.

.venv/bin/python scripts/checkin.py --port 8901           # learn once, then never again
.venv/bin/python scripts/checkin.py --port 8901 --record  # re-learn, ignoring the saved macro

1. Already done. Read the page. If today's check-in is already on it, stop — no click, no model call, nothing to undo.

2. Replay. Run the macro the first run recorded: zero model calls, a few hundred milliseconds, and it refuses rather than guessing when the page has moved on. This is the stage that runs on every ordinary day, and it needs no key at all.

3. Explore. Only when there is no macro, or the saved one no longer matches. The decision model works the page out, and what it did is recorded as a macro so stage 2 takes over tomorrow. Its path is only saved when the page proves it worked.

Pinning --port matters for the demo: a page's origin includes its port, so a second run on a different port is a different site to the browser, with an empty localStorage and no memory of having checked in.

For a real site:

.venv/bin/python scripts/checkin.py --url https://example.com/rewards \
    --goal "Click the daily check-in button" --expect "已签到"
.venv/bin/python scripts/checkin.py --url https://example.com/rewards --replay-only

The goal and the proof are separate arguments on purpose. --goal is what the model is asked to do; --expect is the text the page must show afterwards, checked by code — so a run is judged by the page, never by the model's summary of its own work. Sign in once with --headed --wait 120; the browser profile persists, so later runs reuse the session. --replay-only never calls the model, which is the flag you want in a cron job.

Layout

jev_ultrafast_mcp/
  js/observer.js   in-page observer: stable refs, shadow/frame traversal, verify/resolve
  cdp.py           synchronous CDP client + Chrome launch (no wrapper library)
  browser.py       sessions, guarded execution, op dispatch, macro recording
  observe.py       element model, compact renderer, delta computation
  macros.py        semantic descriptors, scored re-resolution, storage
  assertions.py    deterministic checks
  policy.py        optional TypeSafe turbo policy (speculative fan-out)
  safety.py        domain envelope, redaction, confirmation rules
  config.py        environment-driven configuration
  server.py        the MCP surface
scripts/
  install.py       writes the right config for each MCP client on this machine
  smoke.py         end-to-end proof against a real browser
  mcp_check.py     drives the server over real stdio MCP
  live_check.py    the same, against real websites (needs the network)
  turbo_check.py   lets the decision model drive a real browser (needs a key)
  checkin.py       a real check-in: learn once with the model, then replay for free
examples/
  checkin.html     the daily-button page checkin.py drives
assets/
  social-preview.png      the card GitHub shows when this repository is shared
  make_social_preview.py  renders it, so the words on it are placed rather than generated
llms.txt                  what this server is, for agents that read before recommending it
server.json               the MCP registry entry (lands once the package is on PyPI)

Development

.venv/bin/pip install -e ".[dev]"
.venv/bin/ruff check .
.venv/bin/python -m pytest -q
.venv/bin/python scripts/smoke.py        # 58 checks, real browser
.venv/bin/python scripts/mcp_check.py    # 17 checks, real stdio MCP

All of it runs in CI on Python 3.10, 3.12 and 3.13 against headless Chrome. Read CONTRIBUTING.md before changing how targets are resolved — that logic is the whole point of the project.

See also

• browser-use/jev-ultrafast — the project the optional turbo path is a port of: one typed question per step, answered by a decision model.
• TypeSafe — the typed-question decision API behind browser_goal.
• Model Context Protocol — the protocol this server speaks.
• The official MCP registry — where clients go looking for servers like this one.
• Chrome DevTools Protocol — what it drives the browser with, through no wrapper library.

License

MIT — see LICENSE.