npm • PyPI • Folate • Apache 2.0
A drop-in replacement for agent-browser that pairs ultra-lightweight DOM summaries with local Folate neural DOM grounding—cutting agent token costs by up to 90%+ per step.
Instead of dumping full accessibility trees or hundreds of @eN element references on every snapshot, Frowse returns a compact semantic summary of the page—just enough for an AI agent to craft a precise natural-language query for Folate, which traverses the live DOM locally on CPU (10–25 ms, < 250 MB RAM, $0 API cost) and resolves the exact CSS/XPath selector.
Overview
| Capability | agent-browser (Full Refs) |
Frowse (snapshot + Folate) |
|---|---|---|
| Snapshot Output | Full accessibility tree / @e1..@eN list |
Deduplicated region summary (~100–300 tokens) |
| Repeated Grid / List Items | Every item emitted individually | Collapsed ("Add to cart" (x100), +92 more) |
| Element Targeting | @eN refs or manual CSS/XPath |
Natural-language Folate prompts, CSS, XPath, or @eN |
| Selector Grounding Cost | Agent context tokens | $0 (100% local CPU INT8 ONNX via folate) |
| CLI Compatibility | agent-browser |
Drop-in frowse & agent-browser binaries |
Key Capabilities
- Ultra-Lightweight DOM Summary — Groups visible interactive elements by semantic landmark (
dialog,header,nav,form,main,aside,footer), preserves input types, labels, placeholders, and select options, and deduplicates repeated actions to keep token usage minimal. - Natural-Language Folate Grounding — Pass plain descriptions directly to
click,fill,type,select,hover,check,get, orwait(e.g.frowse click "Run Search button"), or inspect ranked matches withfrowse query. - Persistent Session Daemon — Keeps headless Chromium and the quantized
folate-int8ONNX model warm in memory across CLI calls via per-session Unix sockets (--session <name>). - Drop-In
agent-browserCompatibility — Exposes bothfrowseandagent-browserCLI binaries, supports--json,--headed,--cdp,--refs(@e1,@e2), tabs, iframes, cookies, storage, dialogs, and screenshots.
Quickstart
Installation
Install globally via npm:
npm install -g frowse
Or run directly with npx:
npx frowse open https://example.com
Note: Frowse requires Python 3.10+ on
PATH. On first run, the CLI automatically provisions an isolated environment in~/.frowse/venviffrowseis not already installed.
Or install from source:
git clone https://github.com/figranium/frowse
cd frowse
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
Usage
# 1. Open a page (persistent session stays alive in the background)
frowse open https://example.com
# 2. Get a compact, Folate-ready DOM summary
frowse snapshot
# 3. Act using natural-language Folate prompts (or CSS / XPath / @eN)
frowse fill "Search packages input" "folate"
frowse select "Country dropdown" "Canada"
frowse click "Run Search button"
# 4. Read back state and close the session
frowse get text "#out"
frowse close
You can also query Folate directly without performing an action:
frowse query "Sign in link in header" --top-k 3
And if you need legacy @eN refs for an existing agent-browser workflow:
frowse snapshot --refs
frowse click @e1
Commands
| Command | Description |
|---|---|
open <url> |
Launch browser & navigate (aliases: goto, navigate) |
snapshot / summary |
Lightweight Folate-ready DOM summary (--section, --max-chars, --refs) |
query <prompt> / folate |
Query Folate directly for ranked CSS/XPath selectors (--top-k N) |
click <sel|prompt> |
Click element (--button, --modifiers, --new-tab) |
fill <sel|prompt> <text> |
Clear and fill input |
type <sel|prompt> <text> |
Type into element (--clear, --delay <ms>) |
select <sel|prompt> <val> |
Select dropdown option by value or label |
get <prop> [sel|prompt] |
Get text, html, value, attr, title, url, count, box, styles |
is <state> <sel|prompt> |
Check visible, enabled, or checked |
wait <target> |
Wait for selector/prompt, ms, --text, --url, --load, or --fn |
tab [new|close|<id>] |
Manage tabs (--label <name>) |
frame <sel|@ref|main> |
Switch iframe context |
screenshot [path] / pdf <path> |
Capture screenshot (--full) or PDF |
eval <js> |
Execute JavaScript in the active page/frame |
close [--all] |
Close active session (or all sessions) |
Global flags: --json, --session <name>, --headed, --timeout <ms>, --cdp <port|url>, --viewport <w>x<h>, --model-dir <path>.
Benchmarks
Independent subagent runs (o200k_base CLI output tokens, no shared context) comparing Frowse (0.1.0 with bundled frowse & frowse-extract skills) against agent-browser (0.43.0) across three real-world workflows:
| Workflow | Frowse (o200k tokens) |
agent-browser (o200k tokens) |
Token Reduction |
|---|---|---|---|
| Google Store Pixel Ecosystem (< $1,000) | 15,154 (64 steps) | 59,290 (41 steps) | 74.4% (3.9× lower) |
| Hacker News Top Stories + Comments | 761 (7 steps) | 8,457 (5 steps) | 91.0% (11.1× lower) |
crates.io Package Discovery (ort) |
3,096 (29 steps) | 6,773 (12 steps) | 54.3% (2.2× lower) |
| Total (All 3 Workflows) | 19,011 | 74,520 | 74.5% (3.9× lower) |
Notes: Token counts measure exact
o200k_baseCLI stdout/stderr during browser execution (21,913total for Frowse if also including initialfrowse skills showreads). On all three workflows, both independent agents completed the task end-to-end and arrived at identical verified answers (the same 4-device $977.98 Google Pixel ecosystem on Google Store, the same top 3 stories + top comment on Hacker News, and the sameortcrate metadata on crates.io).
Benchmark prompts
-
Google Store Pixel Ecosystem (< $1,000):
On the Google Store (
https://store.google.com/us/?hl=en-US), find a full Google Pixel ecosystem (Pixel phone, Pixel Watch, and Pixel Buds — plus any other Google Pixel device if it fits) for under $1,000 total (pre-tax), with ONLY: (1) Brand new items (full outright purchase price, no trade-in), (2) Google-branded items only, (3) Sold & fulfilled directly by Google Store. Verify each item's exact configuration, outright price, and URL on its product/configurator page. -
Hacker News Top Stories + Comments:
- Open
https://news.ycombinator.com/ - Find the top 3 stories with at least 50 points (extract title, points, and comment count).
- Navigate to the comments page of the highest-scoring story among those 3, and extract the top-level first comment's author and a brief summary/excerpt of the comment.
- Open
-
crates.io Package Discovery (
ort):- Open
https://crates.io/ - Search for
ortusing the search box. - Open the
ortcrate page from the search results. - Extract and report the latest version, release date/time, license, and repository URL (or Rust version/MSRV if shown).
- Open
Bundled Agent Skills
Frowse bundles four lightweight, bloat-free SKILL.md guides (frowse, frowse-extract, frowse-auth, frowse-debug) covering Folate prompting, bulk extraction handoff, auth/sessions, and debugging:
frowse skills list # List bundled skills
frowse skills show frowse # Print Folate prompting skill
frowse skills install --target claude # Install into .claude/skills/ (or agents, codex, cursor)
License
Frowse is released under the Apache 2.0 License.
Metadata
Release files for frowse 0.1.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| frowse-0.1.1.tar.gz | 36.7 kB | With provenance | Details for frowse-0.1.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| frowse-0.1.1-py3-none-any.whl | Python 3 | none | any | Details for frowse-0.1.1-py3-none-any.whl |
Total release size: 69.6 kB
Release files / frowse-0.1.1.tar.gz
| Download URL | frowse-0.1.1.tar.gz |
|---|---|
| Size | 36.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
33959af2c694e3fb50532d8bf09778ec48226084499d494a8f1dbd63afdcd51b
|
|
BLAKE2b-256 checksum How to use checksums |
70f447bb6cc2f11fcbecb860d887a1abfc4385a19b5961a04c68ff6c8b829416
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logRelease files / frowse-0.1.1-py3-none-any.whl
| Download URL | frowse-0.1.1-py3-none-any.whl |
|---|---|
| Size | 32.9 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
7a00f668ee0fb6a7c77b6e8c6be46826650efe0d7a90c013ecbe97c6681a3528
|
|
BLAKE2b-256 checksum How to use checksums |
ea3f3fd74307339c8cf2ec2b6808c14ba58f5a7cfe44dd07004e50b9d2d75214
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logSecurity for frowse 0.1.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for frowse-0.1.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.