Run a real LLMin your browser. No install.
runs on your device · 0 servers · no install
Run a real LLM
in your browser. No install.AgentOp runs open models locally with WebGPU — write Python tools, test instantly, and export the whole agent as a single standalone HTML file. Your prompts never leave the tab.
$ start building → Browse the marketplace
✓ 100% on-device ✓ Cloud optional ✓ Works offline
NEWGenerate a template with AI — describe your idea, we build it →
models — 100% in your browser hover for details
🦙Llama 3.21–3Blocal›
🌐Qwen 30.6–8Blocal›
🔬Phi 4 Mini3.8Blocal›
💎Gemma 4E2B–12Blocal›
🔍DeepSeek R11.5–8Blocal›
🛠️Hermes 33Blocal›
GGUF · Q4\_K\_Mcached in browser · offline-ready
🔒
PrivateData never leaves your device
⚡
FastGPU-accelerated with WebGPU
🌐
UniversalRuns in any modern browser
📂
OpenOpen models, no lock-in
// one file, infinite reach
Share an agent as a single file. They just open it.
Export bundles the model loader, your Python, the UI and encryption into one HTML file. Send it to anyone — it runs on their machine, offline, with nothing to install on their end. It's yours, and it doesn't expire.
HTML
research-assistant.html
self-contained · 1 file · 0 dependencies
──▶ goes anywhere · recipient installs nothing
📧
#### Email it
Attach the file. They double-click and it just runs — no link to a server.
opens offline
🌐
#### Embed on a site
Drop it in an iframe or host it as a static page. No backend to deploy.
static hosting
💾
#### USB / local drive
Runs straight from file:// — perfect for air-gapped machines.
no network
🛒
#### Publish & fork
List it on the marketplace so others can run it or remix their own.
community-ready
Every shared agent is the whole app in one file — model loader, Python runtime and UI included.
// see it run
Describe an agent, watch AI build it, then run it locally — no install.
Your browser does not support the video tag.
// how it works
From idea to a shippable agent in minutes.
No backend to deploy. No keys to wire up before you can test. Open a template, edit Python, run it right where you are.
01
Pick a template
Start from a curated agent — chat, data analysis, code review — with prompts and tools already wired.
02
Edit & run locally
Write Python that runs in-browser via Pyodide. Test against a local model instantly — no API key required.
03
Export one file
Download a standalone HTML bundle. It runs offline, embeds anywhere, or publishes to the marketplace.
// ✨ ai template generator
Describe it. AI builds the template.
Not sure where to start? Describe your agent in plain English. Your own API key goes straight from your browser to the LLM — AgentOp only sees the finished template.
01
💬
#### Describe your idea
"A research assistant that summarizes academic papers and extracts key findings" is enough to start.
→
02
⚡
#### AI writes the code
Bring your OpenAI or Anthropic key. The complete template — HTML, Python tools, prompts and UI — is generated in seconds, entirely in your browser.
→
03
🚀
#### Review and publish
The draft opens in the full editor. Tweak the code, preview it, then hit Publish to share it with the community.
✓ Your API key stays in your browser. It makes one call to the LLM — it is never sent to AgentOp's servers.
// build with python
Real Python tools, running in the tab.
Define tool schemas and handlers in Python and AgentOp executes them in-browser via Pyodide. Grammar-constrained (GBNF) function calling behaves identically whether you're on a local model or a cloud one.
- ✓ Pyodide 0.29 runtimeThe full CPython scientific stack, no install
- ✓ Unified tool callingSame schema works on every model family
- ✓ Test as you typeRun against a local model with one click
- ✓ OpenAI & Anthropic built inOr add a custom connector
- ✓ Bring-your-own-keyEncrypted in your browser, never stored
- ✓
agent.pylocal
\# tools your agent can call
def get\_tool\_schemas():
return \[{
"name": "search\_docs",
"description": "Search local files",
}\]
def handle\_tool\_call(name, args):
if name == "search\_docs":
return search(args\["query"\])
// cloud when you want it
Local by default. Frontier models on demand.
Need GPT-4-class output for a hard task? Flip a provider on and bring your own key. Keys are encrypted client-side with AES-256 and never touch our servers — so you stay in control of both your data and your bill.
providers
◐ Local · on-device
AI OpenAI byo key
An Anthropic byo key
// the difference
A browser tab vs. a hosted chatbot.
| | AgentOp (local) | Hosted chatbot | | ----------------------- | --------------------------- | ------------------- | | Where inference runs | ✓On your device | On someone's server | | Install / account | ✓None — open a tab | App or sign-up | | Works offline | ✓Yes, after first load | No | | Your prompts & data | ✓Never leave the page | Sent to the cloud | | Cost to run locally | ✓Free | Per-token billing | | What you walk away with | ✓A file you own — no expiry | A subscription |
// recent agents
🤖 ATS Resume Optimizer Agent 🤖 Contract Plain-Language Explainer Agent 🤖 Data Analysis Agent Browse all →
// available templates
📄 Data Analysis Agent Template
📄 Contract Plain-Language Explainer
📄 Resume Optimizer
0
Servers in the loop
// who it's for
Built for people who don't want to hand their data to AI companies.
⌨️
#### Developers
Prototype agents in Python without standing up any infra.
🔒
#### Privacy-first teams
Keep sensitive docs on-device — nothing is sent anywhere.
🎓
#### Educators
Hand out a single HTML file that just runs, offline.
🧪
#### Tinkerers
Fork, remix and break things with zero setup cost.
// faq
The questions everyone asks first.
#### Q.Is it really running locally?
Yes. The model downloads once, caches in your browser, and inference runs on your CPU/GPU. Open your network tab — it's empty during a chat.
#### Q.Do I need a GPU?
No. WebGPU accelerates things if present, but everything falls back to running on your CPU. Smaller models run fine on a laptop.
#### Q.What about my data and prompts?
They never leave the page when you use a local model. There's no server to send them to.
#### Q.Can I still use GPT-4 or Claude?
Yes — bring your own key. It's encrypted client-side with AES-256 and never stored by us.
#### Q.What does sharing a file give the recipient?
The whole app. They open one HTML file and it runs on their machine, offline — no install, no account.
#### Q.How big are the downloads?
From \~0.5GB (Qwen 0.6B) to several GB for larger models. They cache, so it's a one-time cost per model.
// $ npx create-agent
Build a private AI agent today.
Free to start. No credit card. Test locally, then export a file you actually own.
🦙
Llama 3.2
Meta
ctx download
▹