grip

Token-efficient, CDP-native browser SDK for AI agents.

Built directly on Chrome DevTools Protocol — no Playwright, no Puppeteer, no wrapper overhead.

pip install grip-browser

---

What is Grip?

Grip is a CDP-native browser SDK for AI agents that turns a web page into a ~2,000-token semantic snapshot instead of ~78,000 tokens of raw HTML. It runs on the Chrome DevTools Protocol directly — no Playwright, no Puppeteer, no wrapper binary.

Why Grip

Agents don't need the DOM. They need to know what's on the page and what they can act on. Grip sends the model only the interactive elements and visible text — structured, indexed, and fuzzy-matchable.

Measured across 8 real pages (Wikipedia, GitHub, react.dev, BBC, Hacker News, Python docs, arXiv, example.com): median 77,588 tokens of raw HTML becomes 2,018 tokens of grip snapshot — a 19x reduction. Per-page it ranges from 3x on a page that is already tiny to 95x on a heavy SPA.

That 19x is against raw HTML, which is the right comparison if your agent would otherwise put the DOM in the prompt. Against naively tag-stripped text — what a retrieval API sends a model — the reduction is about 1.4x, because most of what grip removes is markup rather than words. Both numbers are measured; use whichever matches what you would otherwise send. Method and data: evaluation/.

Grip vs Playwright MCP vs Puppeteer

	Playwright MCP	Puppeteer	Grip
Tokens per snapshot	not measured	not measured	~2,000 median
Built on	Playwright	Chromium binary API	pure CDP
Shadow DOM traversal	partial	no	full
Fuzzy element match (no selectors)	no	no	yes
Typed error recovery	no	no	yes
Prompt-injection guard	no	no	yes

grip's token figure is measured across 8 real pages (median; 3x–95x reduction vs raw HTML depending on the page). The Playwright MCP and Puppeteer columns are feature comparisons — their token figures are not independently measured here, so they are left blank rather than guessed.

Honest caveat: Playwright and Puppeteer are broader general-purpose automation frameworks with huge ecosystems and cross-browser support. Grip is narrower on purpose — it does one thing (feed an LLM the smallest useful view of a page) and does not try to replace them for human-driven E2E testing.

When to use Grip

• You're building an autonomous or semi-autonomous agent that browses the web and you're paying per token.
• Your agent loop is blowing its context window on raw HTML or screenshots.
• You want typed, recoverable errors (CAPTCHA_REQUIRED, RATE_LIMITED, ELEMENT_STALE) instead of parsing exception strings.
• You need shadow DOM / web-component pages handled without special-casing.

When not to use Grip

• You need cross-browser (Firefox/WebKit) human E2E test coverage — use Playwright.
• Your task is a fixed, deterministic scrape with known selectors and no LLM in the loop — a plain scraper is simpler.

FAQ

Is Grip a Playwright wrapper? No. Grip talks to Chrome over the DevTools Protocol directly. There is no Playwright or Puppeteer dependency underneath.

How does it cut tokens? It sends the model only interactive elements (inputs, buttons, links) and visible text, indexed for fuzzy matching — not the full HTML tree, not a screenshot. A trivial page like example.com comes out at ~50 tokens; a Wikipedia article at ~7,000, down from ~157,000 raw.

Which LLMs does it work with? Anthropic and OpenAI adapters ship in the box; any model works via the LLMAdapter protocol.

Does it handle CAPTCHAs / bot blocks? It detects and classifies them (page.detect_challenge()), and returns a typed error with a suggested recovery action (escalate, backoff, rotate). page.solve_challenge() attempts checkbox, Turnstile and slider stages in-process and only reports success it can verify; image-grid and text challenges come back to your model with a screenshot. No third-party solving service is used, and success rates are unmeasured — see Challenges and automation tells.

What do I need installed? Python 3.11+ and Chrome or Chromium. Grip finds Chrome automatically, and falls back to the Chrome for Testing build that Playwright or Puppeteer already downloaded if no system Chrome is present. Set CHROME_EXECUTABLE to override.

---

The problem

Most browser tools give AI agents raw HTML or screenshots. Raw HTML on a real page runs tens of thousands of tokens — measured median 77,588 across 8 popular sites, and 157,089 for a single Wikipedia article. Screenshots are ~3,000. Both burn through context windows fast and slow your agent down.

What grip does instead

grip gives your agent a semantic summary of what's on the page — just the interactive elements and visible text, structured for LLM consumption:

PAGE: Amazon.com
URL: https://www.amazon.com/

INTERACTIVE:
  [inp:0] "search here" (placeholder)
  [btn:1] "Go"
  [btn:2] "Sign in"
  [lnk:3] "Returns & Orders"

CONTENT:
  Delivering to New York — Shop deals in...

~2,000 tokens per snapshot, median. The example above is example.com, the smallest page on the web at ~50 tokens. A Wikipedia article is ~7,000 — against ~157,000 raw.

---

Quick start

import asyncio
from grip import Browser

async def main():
    async with Browser(headless=True) as browser:
        page = await browser.open("https://news.ycombinator.com")
        snapshot = await page.snapshot()

        print(snapshot.text_content)      # readable page text
        print(snapshot.elements)          # interactive elements only
        print(snapshot.tokens_estimated)  # ~50 for this page; ~2,000 median on real pages

asyncio.run(main())

Full agent loop

async with Browser(headless=True) as browser:
    page = await browser.open("https://amazon.com")
    await page.snapshot()               # build element index

    await page.type("search", "blue sneakers")
    await page.click("Go")              # fuzzy match — no selectors needed

    await page.snapshot()               # re-index after navigation
    doc = await page.read()             # prose, citable blocks, no nav chrome

    shot = await page.screenshot()      # JPEG, ~800 tokens for vision models
    shot.save("result.jpg")

Concurrent pages

Every open() gets its own tab and its own CDP connection, so pages are independent and can be driven in parallel:

async with Browser(headless=True) as browser:
    urls = ["https://example.com", "https://example.org", "https://example.net"]
    pages = await asyncio.gather(*(browser.open(u) for u in urls))
    snapshots = await asyncio.gather(*(p.snapshot() for p in pages))

    for snap in snapshots:
        print(snap.url, snap.tokens_estimated)

    for page in pages:
        await page.close()          # closes the tab; browser.close() also closes any left open

page.goto(url) navigates an existing tab in place. There is no built-in concurrency limit — wrap in an asyncio.Semaphore if you need one, since the safe ceiling depends on your machine rather than on grip.

Read mode

snapshot() answers "what can I click here". read() answers "what does this page say" — main content isolated, navigation and footer chrome dropped, and every block carrying the heading trail above it so a claim can be cited back to a location.

async with Browser(headless=True) as browser:
    page = await browser.open("https://docs.python.org/3/library/asyncio-task.html")
    doc = await page.read()

    print(doc.outline())          # heading map of the page
    for block in doc.blocks:
        print(block.citation, block.text[:60])
        # [12] Coroutines and tasks › Coroutines   Source code: Lib/asyncio/...

read(max_chars=N) truncates by dropping whole blocks, never mid-sentence. The default is no limit — deciding which parts of a page matter is ranking, and that belongs to the caller.

Challenges and automation tells

grip detects checkbox, Turnstile, slider, image-grid, text and invisible challenges from the page's DOM and frame URLs, and classifies them without a network call (page.detect_challenge()). Detection is tested against real widget markup.

page.solve_challenge() implements in-process solve flows for the checkbox, Turnstile and slider stages, using human-shaped pointer motion and no third-party solving API. Each flow reports "solved" only after it verifies the outcome — a response token is present, or the widget has left the page. If neither is true when the timeout expires it returns "timeout", never "solved". Image-grid and text challenges return "needs_vision" with a screenshot for your own model to answer; grip does not ship a classifier. Solve success rates are unmeasured as of 2026-08-10: they depend on IP reputation and provider-side scoring, so any number quoted here without a stated egress would be meaningless.

result = await page.solve_challenge(timeout=30.0)
match result.status:
    case "solved":       ...  # verified: token present or widget gone
    case "needs_vision": ...  # result.screenshot -> your model -> page.click_at(x, y)
    case "unsupported":  ...  # named in result.stage
    case "timeout":      ...  # NOT solved; the challenge is still there
    case "none":         ...

Human-shaped input is available on its own: page.click_at(x, y, human=True) and page.drag(start, end) travel a curved, eased Bézier path with a randomized press dwell. Straight-line constant-velocity motion is the clearest synthetic-input tell. page.click(desc, human=True) uses that path instead of the JS click: it re-resolves the element first, so it still raises ELEMENT_STALE on a stale handle and clicks the element's live position rather than the one the snapshot recorded. The default stays the JS path — it is faster and works headless — and human=True is for challenge flows.

Chrome under CDP sets navigator.webdriver and puts HeadlessChrome in the user agent. Browser(stealth=True) removes both. It is off by default because grip is a general-purpose SDK and silently masking automation would surprise anyone using it for ordinary testing. Whether that flag helps is unmeasured. A competitor measured the equivalent page-world approach against live reCAPTCHA and found it made detection easier, not harder, so grip does not claim a benefit it has not observed. evaluation/stealth_measurement.py is the script that settles it; it needs a host with outbound network, which the development sandbox is not:

.venv/bin/python -m evaluation.stealth_measurement

grip does not hide that it is automation at the network layer. TLS/JA3 fingerprints, and full headless fingerprint parity, live below the Chrome DevTools Protocol and cannot be reached from a Python client driving stock Chromium. If a site blocks you on IP reputation or TLS fingerprint, no flag in this library will change that — that is an egress problem, and the answer is a residential or mobile proxy, which grip supports via proxy=.

With an LLM (autonomous mode)

from grip import Browser
from grip.adapters.anthropic import AnthropicAdapter

llm = AnthropicAdapter(api_key="sk-ant-...")

async with Browser(llm=llm, headless=True) as browser:
    result = await browser.run(
        goal="Find the cheapest blue sneakers under $80",
        url="https://amazon.com"
    )
    print(result.data)
    print(f"Used {result.tokens} tokens")

grip handles the snapshot → decide → act loop automatically. You just provide the goal.

Snapshot delta

Inside the run loop, grip sends the model a full snapshot on the first turn and a delta after that — only the elements and content that changed. On a 5-turn navigate/click loop over a 28-element page with a 25-result body, the per-turn page payload went from 628 tokens to 42, a 75% cut. Method: build_delta + format_delta against Summarizer.format, counted with tiktoken cl100k_base.

The larger effect is on the transcript rather than on any single turn. Superseded page states are not re-sent, so cumulative prompt cost over a run grows with the number of turns rather than with their square. The first snapshot stays resident in the user message alongside the goal, so roughly 628 tokens of that 5-turn run is a constant floor, not a growth term.

---

Why not Playwright or Puppeteer?

	Playwright MCP	Puppeteer	grip
Tokens per snapshot	not measured	not measured	~2,000 median
Shadow DOM traversal	Partial	No	Full
Prompt injection guard	No	No	Yes
Typed error recovery	No	No	Yes
Element staleness detection	No	No	Yes
Pure CDP (no binary bloat)	No	No	Yes
Screenshot token tracking	No	No	Yes

---

Structured errors

Every error comes back as a typed BrowserError — not a bare string — so your agent can make decisions:

from grip import GripError
from grip.errors.types import ErrorType, RecoveryAction

try:
    await page.click("checkout")
except GripError as e:
    match e.error.type:
        case ErrorType.CAPTCHA_REQUIRED:
            # recovery: ESCALATE_TO_HUMAN or VISION_FALLBACK
            await escalate(e.error.message)
        case ErrorType.RATE_LIMITED:
            # recovery: EXPONENTIAL_BACKOFF + RETRY
            await asyncio.sleep(30)
            await page.click("checkout")
        case ErrorType.AUTH_REQUIRED:
            # recovery: ESCALATE_TO_HUMAN
            raise NeedsLogin(e.error.message)
        case ErrorType.ELEMENT_STALE:
            # recovery: RE_SNAPSHOT + RETRY
            await page.snapshot()
            await page.click("checkout")

Full error taxonomy

Type	When	Suggested recovery
ELEMENT_NOT_FOUND	fuzzy match failed	re-snapshot, retry with different description
ELEMENT_STALE	element moved after navigation	re-snapshot
ANTI_BOT_BLOCK	Cloudflare, DDoS-Guard, 403	rotate identity
CAPTCHA_REQUIRED	CAPTCHA challenge page	escalate to human
RATE_LIMITED	429 Too Many Requests	exponential backoff
AUTH_REQUIRED	login wall	escalate to human
ZERO_RESULTS	page loaded, no matching content	retry, broaden query
NETWORK_TIMEOUT	navigation timed out	exponential backoff
NAVIGATION_FAILED	blank page / bad URL	retry

---

Shadow DOM

grip traverses shadow DOM trees automatically. Web components, Chrome extensions, custom elements — all discovered in the same snapshot:

snapshot = await page.snapshot()
shadow_elements = [el for el in snapshot.elements if el.in_shadow_dom]

---

Trace

Every action is recorded with timing and token cost:

async with Browser() as browser:
    page = await browser.open("https://example.com")
    await page.snapshot()
    await page.click("Learn more")
    await page.screenshot()

print(browser.trace.total_tokens)   # total tokens used
browser.trace.to_jsonl("audit.jsonl")  # machine-readable audit log

---

LLM adapters

grip ships with OpenAI and Anthropic adapters out of the box:

from grip.adapters.openai import OpenAIAdapter
from grip.adapters.anthropic import AnthropicAdapter

llm = OpenAIAdapter(api_key="sk-...")         # gpt-4o, gpt-4-turbo, etc.
llm = AnthropicAdapter(api_key="sk-ant-...")  # claude-opus-4-7, etc.

Or bring your own by implementing the LLMAdapter protocol:

from grip.adapters.base import LLMAdapter, LLMResponse

class MyAdapter:
    async def complete(self, messages, tools) -> LLMResponse:
        ...

---

Requirements

• Python 3.11+
• Google Chrome (or Chromium) installed

grip finds Chrome automatically. Override with CHROME_EXECUTABLE env var.

---

Install

pip install grip-browser

# with OpenAI support
pip install grip-browser[openai]

# with Anthropic support
pip install grip-browser[anthropic]

---

Measured numbers

Everything in this table was measured on this branch. Anything not in it is not claimed: cold-start time, memory, requests per second, challenge solve rates and tokens against another tool are all unmeasured, and quoting them would be a guess. The snapshot-size figures live in Why Grip with their own method note.

	Measured	How
Per-turn page payload, delta on	628 → 42 tokens (75% cut)	5-turn navigate/click loop, 28-element page, 25-result body; build_delta + format_delta vs Summarizer.format, tiktoken cl100k_base
Cumulative prompt cost over a run	grows with turns, not turns²	superseded page states are not re-sent; first snapshot is a ~628-token constant, not a growth term
Unit tests	249 pass	pytest tests/unit
gripsearch tests	33 pass	pytest in gripsearch/
Integration tests	74 pass	real Chrome, live network
Unit coverage	84.18%	unit tests only; CI fails below 80
Lint	ruff 83 → 0	both gates previously passed vacuously because neither was configured
Types	mypy --strict 35 → 0	as above
example.com, live	open 0.80s, snapshot 0.01s, 1 element, 50 tokens	headless Chrome, single page
Local file fixture	open 0.61s, snapshot 0.01s	file:// page, no network
Chrome profile directories stranded	0	across a 57-minute full-suite run

Test and lint counts are for this branch and will move. Re-run them rather than trusting the table if the number matters to you.

---

Contributing

Contributions are welcome. See CONTRIBUTING.md for dev setup, running tests, and lint/type-check commands. Please also read the Code of Conduct. Found a security issue? See SECURITY.md instead of opening a public issue.

---

License

MIT — see LICENSE.