grip — Complete Design

A Python SDK for AI agents to interact with the web. Token-efficient, secure, built on CDP. Review each section and confirm in the terminal.

What it is
grip
A Python SDK that gives AI agents a precise, token-efficient, secure way to read and interact with any website.
pip install grip-browser
Core API — what a developer writes
Simple — let grip handle everything
from grip import Browser

browser = Browser(llm=my_llm)

result = await browser.run(
  "Find cheapest blue sneakers",
  url="amazon.com"
)

print(result.data) # structured
print(result.trace) # full trace
print(result.tokens) # always tracked
Precise — full control for your agent loop
page = await browser.open("amazon.com")

# compact — not raw HTML
state = await page.snapshot()

# semantic — not CSS selectors
await page.click("search bar")
await page.type("blue sneakers")

data = await page.extract({
  "results": "list[Product]"
})
Structured Errors — what the agent receives when something breaks
# Instead of: "Error: element not found"
# grip returns:

BrowserError(
  type=ErrorType.ELEMENT_STALE,
  message="Button ref stale after React re-render",
  confidence=0.94,
  recovery=[
    RecoveryAction.RE_SNAPSHOT, # take fresh snapshot
    RecoveryAction.RETRY # retry the action
  ]
)
What ships in V1 vs later
V1 — Ships first
CDP engine (Chrome/Chromium launch + connect)
Compression layer (semantic summaries, element cache, snapshot versioning)
Security sanitizer (injection guard, hidden filter)
Typed error recovery (6 error types)
Full Shadow DOM access
Primitive API: snapshot, click, type, extract, observe
High-level: run(goal, llm)
Observability trace (token counts, action log)
LLM-agnostic (works with any model)
V2 — After validation
Vision fallback for canvas apps (Figma, Sheets)
Cloud browser support (Browserbase, Steel)
Advanced bot detection evasion
TypeScript SDK
WebMCP support (when Chrome ships it broadly)
Benchmark suite (realistic network faults)
Does this look right? If yes, I'll write the full spec document and we move to implementation planning.