One agent for the open web
Checkout, a job application, a multi-step wizard, reading one page to write on another — no integration, no scripts.
Open source · MIT · Bring your own
key
Your browser,
Your browser,
driven by AI.
OpenSidebar is an open-source Chrome extension that puts an autonomous agent in your side panel — describe a task in plain English and it sees the page, clicks, types, and carries multi-step work across tabs to done.
Bring your own API key. No subscription, no telemetry, no backend of ours — the extension runs in your browser and talks only to the model providers you configure.
Quick start
From clone to first task
The Chrome Web Store listing is on the way. Until then, build from source and load the unpacked extension — the steps below are the entire process.
Chrome Web Store
One-click install
A signed listing so anyone can add OpenSidebar straight from Chrome.
Coming soonFrom source
Load the unpacked build
# clone & build git clone https://github.com/krisshkodrani/OpenSidebar cd OpenSidebar corepack enable corepack pnpm install corepack pnpm run dist # then in Chrome chrome://extensions → Developer mode → Load unpacked → select dist/
Requires Node.js 22+ and a supported provider API key.
Recorded live
See it actually do the work
Every clip is a real session the agent drove end to end, captured as it ran. Scenes are trimmed and captioned for length — nothing is mocked up or re-enacted.
Operates enterprise apps, live
The same agent drives ServiceNow — creating incidents, ordering from the catalog, filtering lists. One agent, every workflow.
Watch Mode keeps an eye out
Leave it watching a page and it re-checks every few seconds — in this run it flagged “back in stock” within seconds of the flip.
Your provider, your models
Paste a key, pick your models, go. Fireworks, OpenRouter, Moonshot, or Xiaomi — nothing hardcoded.
Watch the full tour
Three acts — the open web, staying in control, and the ServiceNow finale. Narrated.
Real sessions, unscripted — edited only for pacing and captions. Executor: Kimi K2.7 Code (vision) · Planner: GLM 5.2 · Judge: GPT-OSS 120B · Fireworks AI
Under the hood
See the plan, not just the run.
Every demo above is a recorded run — and every run starts with a plan. The walkthrough page replays four real traces click by click: the prompt, the planner's subtask chain, the executor's actual page actions, and the verifier's verdict on each step, with the model named at every stage.
Four recorded runs: a checkout order · a cross-page inventory lookup · a guarded settings cleanup · a ServiceNow incident
What it does
An agent that sees, plans, acts — and stays in your control
Sees the page like you do
A vision model works from the live screenshot and DOM — reads charts, zooms into fine print, and handles pages that defeat text-only bots.
50+ browser tools
Clicking, typing, and scrolling, plus file upload and download, tab and window management, and structured read-outs of tables, chart labels, and active filters — the agent gets data out of pages, not just actions into them.
Safe by default
Approvals are on by default, and consequential actions — submitting a job application, sending a message — always pause for you. When the agent works from a draft you approved, the final submit is checked field-by-field against that draft; any mismatch stops and shows you the difference before anything is sent.
Plans, executes, verifies
A planner decomposes the task, an executor drives it, and a verifier confirms the result — high-risk completions get a second opinion from a dedicated judge model. Stuck runs escalate to a stronger model instead of burning turns.
Remembers what matters
A local profile you review yourself, per-site skills you record once and reuse, and checkpoints that survive restarts (kept for 24 hours). Sensitive fields are consent-gated per task and encrypted with a key that never leaves your machine.
Watch Mode
Point it at a page and it re-checks every few seconds, comparing what actually rendered — then tells you when a price, a status, or a listing changes.
Speak or type
Add a Groq API key in settings and voice input transcribes your words straight into the composer (Whisper large-v3-turbo) — describe the task out loud and let it run.
Yours to inspect
Every session produces a full-fidelity local trace. Keys live in Chrome storage, and traffic goes from your browser to the providers you configure — your LLM provider, plus Groq if you enable voice — and nowhere else.
Bring your own key
Your provider. Your models. Your data.
Fireworks recommended
OpenRouter
Moonshot / Kimi
Xiaomi MiMo
Hybrid stacks add DeepSeek-, Cerebras-, and Groq-served models. Your API keys stay in Chrome storage; page context goes only to the providers you configure. No analytics, no tracking, no hosted relay, no OpenSidebar servers.
In the open-source repo · developer builds
Every run leaves evidence.
The repo ships a full observability workspace: replay any run as a story — plan, nodes, turns, verification — with the judge model's per-criterion reasoning and cost on high-risk completions, then record whether the outcome was actually right. Traces never leave your machine.
- MIT licensed. Audit every claim on GitHub.
- Nx + pnpm monorepo. Manifest V3, Vitest, staged e2e harness.
- Benchmarks with receipts. Neutral public numbers, published with per-task judge outputs — coming shortly.
