Endoplexity

The numbers · Versus the field · How it works · Get started · Guide

▶ Play the 26-second film, sound on · every number in it is from the benchmark below · an unedited real runMusic: "Inspiring Progress" by Soundrider/Dope, Jamendo, CC BY 3.0

Why Endoplexity

Your Chrome, not a new browser. It works in the profile you already live in: your logins, extensions and password manager. Nothing to move into, nothing to sign in to again.

Your subscription, not an API key. Tasks run through the claude, cursor-agent or codex CLI you already pay for. On top of that plan, a task costs $0.

Text, never screenshots. The model reads the page as a compact outline, then only what changed. In 108 benchmark runs on three vendors' models, not one screenshot reached a model.

Brakes the model cannot release. Irreversible clicks stop for you. A second model from a different company reads each one first. The rules live in local code, never in a prompt.

Learn once, replay free. Save a finished task as a skill and it runs again with no model at all: zero tokens, the same approval gate.

Reflexes that make it faster. Small decision models answer routine questions in milliseconds, so the big model doesn't have to. Runs finish 25% faster with them on.

The numbers

Thirty real tasks on thirty live websites from Online-Mind2Web, frozen before the first run and judged blind. Against Claude in Chrome, Anthropic's own extension, on the same model through the same CLI: only the browser tools differ.

On the 20 tasks both finished, Claude in Chrome took longer, read more and cost more at the median. Every row, the judging and what the numbers don't show: docs/benchmarks.md. Method and raw evidence: bench/METHOD.md.

Versus the field

	Endoplexity	Perplexity Comet	Claude in Chrome	ChatGPT agent	Gemini in Chrome
Drives your own Chrome	✓	separate browser	✓	cloud browser	✓
Model never sees a screenshot	✓	—	—	—	not published
Your choice of model	Claude, Cursor, Codex	Perplexity's	Claude	OpenAI	Gemini
Extra cost on your plan	$0	paid plans	paid Claude plan	Plus: 40 tasks/mo	AI Pro or Ultra, US
Replays a task for free	✓	—	—	—	—
Open source	✓ Apache-2.0	—	—	—	—

Checked 2026-11-07 from each vendor's pages and, for Comet and Claude in Chrome, their shipped extension code. Opera Neon, Browser Use and every source: the full comparison.

Why not just use Comet?

Comet asks you to move into a new browser, sends a JPEG of the page after every action to Perplexity's servers, and runs a model Perplexity picks. Endoplexity stays in your Chrome, sends text only to the model vendor you already use, and keeps its brakes in code on your machine. That last part matters: hidden text on a page has already walked Comet into reading a user's one-time password (Brave), and a single crafted link walked it into leaking email and calendar data (LayerX). No agent is immune to what it reads, so Endoplexity's approvals never depend on the model behaving.

How it works

The CLI you already have is the brain. Endoplexity adds the hands (a local bridge with 18 browser tools) and the face (a side panel that shows every step). The extension never talks to a model, the CLI never touches Chrome, and the safety gate sits in between, where nothing the model says can move it.

Reflexes: Laya and Jev

Most of what an agent decides on a page is multiple choice: which element is the Apply button, is this click irreversible, is this a cookie wall or a login. Reflexes answer those with small decision models instead of a frontier model's turn.

Laya, local and free. Convai's open model, 421M parameters, on your own GPU at about 100 ms a decision. It learns from your runs: npm run reflex:train fine-tunes it on what the big model decided, and keeps the new weights only if they score better.

Jev, hosted and sharp out of the box. TypeSafe AI's decision model, 75% right on our held-out decisions with no training, and it sees pixels: on a bare canvas it placed three shapes at p = 0.99 and clicked each one. About $0.0002 a decision, off by default, opt-in twice.

A reflex earns the right to act: Shadow first, Act only after 50 decisions at 95% agreement. Any reflex may add an approval; none may remove one.

Get started

You need Node 24+, Chrome 114+, and one of the claude, cursor-agent or codex CLIs, logged in. A Claude Pro plan is enough.

git clone https://github.com/Endokelp/Endoplexity.git && cd Endoplexity
npm install
npm run setup   # Windows: start the bridge at every login
npm start       # or start it now, on any OS

Then open chrome://extensions, turn on Developer mode, click Load unpacked and pick the extension/ folder. Open the side panel, type what you want done, press Run.

Files, Cursor and Codex logins, parallel lanes, scheduled runs, the plan lock and the audit export are all in the guide.

Safety, in short

• The bridge listens on 127.0.0.1 only, and only your extension's pinned origin can connect.
• The agent gets browser tools and nothing else: no shell, no web fetch, no files beyond the ones you attach, verified by asking each CLI to reach past its tools.
• Submitting, paying and deleting stop for your approval. Silence denies.
• Files go to a page by a key you configured, never by a path the model picks.
• Nothing the model proposes (a site note, a changed skill) is saved until you approve it.

The full model, including where it stops: the guide. Found a hole? SECURITY.md.

Status

v0.2.0. Windows-first (autostart), portable everywhere else with npm start. 318 tests, no framework, no build step. Known gaps are listed in the guide, not hidden.

Licence

Apache-2.0. See NOTICE for trademark attribution.

Independent project. Not affiliated with, endorsed by, or sponsored by Perplexity AI, Anthropic, Anysphere, OpenAI, Google, Convai or TypeSafe AI.