Browse the web at 17s median speed and a fraction of the cost.
Jev chooses each action from indexed page controls, an LLM plans and reads, and deterministic code guards verification, irreversible gates, and secrets.
Core Philosophy
Picks instead of generating
Most browser agents generate actions from screenshots, creating brittle CSS selectors that fail and rack up retry loops. Fastbrowse breaks this paradigm.
Jev Choice Model
Indexes interactive DOM elements into discrete candidates. A specialized choice model selects from what actually exists on the page—it is physically impossible to click a phantom element.
Verbatim Quoting
Every single claim in the final answer is backed by a verified, verbatim quote from the page. If the fact isn't explicitly present on screen, the agent flags it rather than guessing.
Irreversible Action Gates
Code-level safety guarantees. Payment buttons, delete actions, and outbound emails automatically pause execution at needs_confirmation unless run with --authorize.
Zero-Knowledge Secrets
Models never see passwords or sensitive tokens. Fastbrowse matches origin policies via Bitwarden or environment variables, injecting credentials directly at typing time.
Local or Cloud Browser
Runs headfully or headlessly on your local Chrome profile, or switches seamlessly to Browser Use Cloud via --cloud to bypass aggressive anti-bot captchas.
Embeddable Python SDK
First-class Pydantic schema validation. Return structured data directly from arbitrary web journeys into your backend pipelines with strict step and dollar limits.
Empirical Results
Head-to-head against Browser Use
The same 14 answer tasks (lookups, sign-ins, checkout, Google Flights), three passes each, on identical cloud browsers under identical limits (30 steps, $0.25, 300s).
| Metric | fastbrowse | Browser Use (Hosted) | Delta |
|---|---|---|---|
| Tasks Passed | 40 / 42 | 14 / 42 | +2.9× pass rate |
| Median Execution Time | 17.0s | 27.5s | 1.6× faster |
| Average Cost per Task | $0.011 | $0.400 | 35× cheaper |
| Action Selection Method | Jev choice model on indexed DOM | LLM generates from raw screenshot | Deterministic vs generative |
| Fact Grounding | Verbatim quotes cited with URLs | Unverified synthetic text | 100% cited claims |
Full breakdown, evals runner scripts, and raw traces documented in the evals repository documentation.
System Design
Separation of Planning, Choice & Safety
The task is planned and the start page opened in parallel; each step indexes the page, Jev picks an operation and target, code gates it and acts.
LLM Planning & Reading
Gemini 3.8 Flash & Gemini 3.5 Flash Lite extract info, establish goals, and verify quotes.
Jev Choice Model
Classifies interactive elements and picks the exact control to advance page state.
Deterministic Code
Enforces authorization gates, intercepts cookie banners, and handles vault secrets.
Get Started
Ready in 60 seconds
# Clone and setup with uv
git clone https://github.com/agent-labs-dev/fastbrowse.git && cd fastbrowse
uv sync
cp .env.example .env # add AI_GATEWAY_API_KEY or TYPESAFE_API_KEY, and OPENROUTER_API_KEY
# Run with natural language task and start origin
uv run fastbrowse "What is the title of the top story right now?" \
--start https://news.ycombinator.com/
# Or run with authorization gates and secrets
uv run fastbrowse "Log in as standard_user and add the backpack to the cart." \
--start https://www.saucedemo.com/ \
--secret password=SAUCE_PASSWORD \
--authorize