Skip to content

OpenBrowserHarnessThe agent that lives in your browser

A browser extension for Chrome and Edge. It opens real pages, fills forms, and extracts content — visibly, and under your approval.

INSTALL
  1. 01

    Download openbrowserharness.webstore.zip from Releases and unzip it.

  2. 02

    Open the extensions page at chrome://extensions (Edge: edge://extensions).

  3. 03

    Enable Developer mode, click Load unpacked, and pick the unzipped folder.

BY THE NUMBERS

60+browser tools
8presets & modes
3permission tiers
100%local session log

HOW IT RUNS

One instruction, three journeys

01

You state the goal

Tasks arrive in plain language: research, fill forms, watch a page. Slash commands, @ subagents, and skill capture are one keystroke away.

02

It drives real pages

Humanized input and a visible cursor operate real tabs — multi-tab management, Shadow-DOM piercing, per-keystroke typing — every step approvable and interruptible.

03

Everything is traceable

All model-visible input lands in an append-only session log; the trajectory view replays the turn on a timeline, and logs export for review.

DESIGN NOTES

Visible actions, durable memory

OPERATE

Watch it work

The virtual cursor follows Bézier paths, types per keystroke, and scrolls with inertia. Pages are driven through the browser debugging protocol — never OS-level automation — so every click happens on the real page, visible and interruptible.

The agent calls tabs_open on a Wikipedia article in the side panel and answers with a three-point summary

LEARN

Learned sites, one step to act

A single /learn-site command mechanically learns a site and its same-origin pages and condenses paths, forms, and stable selectors into a site-knowledge plugin. Return to that site and the cheat sheet enters the model context automatically — it operates by known selectors instead of rediscovering the page, faster and cheaper.

The site-knowledge record in the user-plugin roster, toggleable after learning

REVIEW

Every run is traceable

Everything the model sees is recorded in an append-only session log: system prompts, tool calls and results, approval decisions, and every context injection. The trajectory view replays a turn as timeline cards, with input, model, and tool swimlanes showing where the time went — and logs export for review.

The trajectory timeline cards replay user, tool call and result, and assistant steps, with swimlane timing strips above

PRESETS

Eight presets, four working styles

Every preset keeps the full tool set and your model route, encoding only its working discipline as a prompt addendum. New sessions start full-access; the shield knob tightens it anytime.

The scenario-preset dropdown: web research, deal hunting, video ops, form running, page monitoring, dev probing
Standardfull tools

Full tool set: browsing, actions, script checks, site knowledge, plans, subagents

PTCprogram first

Synthesizes multi-step operations into one program run to cut round trips

Minimalminimal calls

Fewest tool calls — one step when one step suffices

Createmake presets

Turns a role, working style, and boundaries into a saveable custom preset

Web researchermulti-source

Deep reads, multi-source cross-checks, cited conclusions

Page sentinelread-only watch

Watches pages on a cadence, read-only, reports only on change

Dev proberead-only probe

For developers: read-only DOM and console probing to pin down page issues

Market watcherscheduled watch

Follows finance creators, distills stock and market signals, daily schedule optional

Hand your browser to an agent you can watch

Download the extension, add your model key, and start with your first instruction.