
Fingerprints from the kernel
Canvas, WebGL, and hardware signals come from a fingerprint build of Chromium. Nothing is injected into pages.
Open-source desktop AI agent
Isolated fingerprints and persistent sessions give every account its own identity. Verified web workflows become Browser Skills, ready to call on demand.
One sentence is enough. Pick a model, a workspace, and whether it asks before each step.
Each environment is a separate identity with its own logins. Piskie reads what each one is for.
It reads the code, edits the file, and runs the tests, showing every step.
Signed in as Operations, it clicks through the admin like a person would. You can watch live and take over.
The steps it took become verified code. Next time the whole task is one call.
How a skill gets made
A scout agent maps the site’s pages, forms, and risks, within ten minutes.
You approve the list of tools and the test scenarios before any code is written.
A builder writes each function against the live site and calls it for real. Logins and CAPTCHAs come to you.
A separate verifier tests every function in a fresh context. Only a full pass is published.
Skills compile in a sandbox: no network calls of their own, no eval, no file system. Only the browser API Piskie gives them.
Chat and schedules
Connect Feishu, WeCom, QQ, or a personal WeChat account. Piskie replies in the same chat with results and screenshots. New senders pair with a one-time code first.
Browser environments

Canvas, WebGL, and hardware signals come from a fingerprint build of Chromium. Nothing is injected into pages.
Each environment keeps its fingerprint and cookies, so sites see the same device each time.
HTTP or SOCKS5 per environment. Timezone, language, and location can follow the proxy’s exit IP.
WebRTC is restricted in every profile, so your real IP stays hidden.
See the agent’s browser live inside Piskie and step in with your own mouse and keyboard.
An environment in use is locked, so two runs never fight over one account.
On your machine
Models and tools
Any OpenAI- or Anthropic-compatible API. A signed catalog covers 780+ models from 15 providers.
Local and remote servers with OAuth. Project servers need your approval to run.
Install from a folder, Git, or zip. Reads Claude Code and OpenAI plugin marketplaces.
Your own ComfyUI workflows, plus Gemini, DashScope, OpenRouter, and Baidu.
Under the hood
At 85% context, Piskie summarizes with the exact prefix the model already saw, so cached tokens stay cached.
Tool schemas share a fixed budget. The rest load on demand through tool search.
Every page, iframe, and worker waits on the DevTools protocol until locale, timezone, and UA are set.
Seeds come from each environment’s ID, aligned to values the kernel reports consistently.
Updates are Ed25519-signed and hash-checked, with rollback protection.
Publishing checks the source still matches the build that passed verification.
FAQ
Is Piskie free?
Free.
Piskie is open source under the MIT license. You only pay your model provider for the API usage you choose.
Which models can I use?
Any compatible API.
Anything OpenAI- or Anthropic-compatible, such as OpenAI, Anthropic, DeepSeek, Gemini, OpenRouter, Qwen, and Zhipu. Each sub-agent can use a different model.
Do I need to install a browser?
No.
On first launch, Piskie downloads and verifies its fingerprint Chromium in the background.
Does it work on Intel Macs?
Not yet.
The macOS build is for Apple silicon. Windows and Linux builds are x64.
Where is my data stored?
On your computer.
Everything lives in ~/.piskie. Model requests go straight to the provider you configure.
Does it run when the app is closed?
No.
Schedules and chat bots run while the app is open. A missed schedule runs once when it’s back.
Open source on GitHub