You already have a DeepSeek, Claude, or ChatGPT tab open. ZeroKey turns that browser
session into a local OpenAI-compatible endpoint your IDE can call — no API keys, no
separate bill. Paste one
fetch(), point your IDE at
localhost:7250, and you're coding in under 2 minutes.
MIT-licensed · open source on GitHub · zero telemetry · runs entirely on your machine
Your browser session is the credential. Nothing to generate, rotate, or leak.
Runs entirely on your machine. No relay servers, no telemetry, no third party in between.
Standard
/v1/chat/completions
— any IDE that speaks OpenAI, speaks ZeroKey.
ZeroKey doesn't sell access — it forwards your own session to your own IDE. Every number below is a fact about the project, not a growth metric.
pnpm start
Every number above is a fact about this codebase — nothing modeled or estimated.
ZeroKey sits in one place: a local process translating OpenAI-shaped requests into the exact calls your browser would make — then streaming the answer straight back.
ZeroKey isn't a fixed pipeline — the tool layer is designed to be extended. MCP servers, custom skills, and per-IDE tool grammar all plug into the same engine.
Tools named
mcp_<server>_<tool>
auto-register as a
$<server>
skill tag — no manual wiring. A built-in Playwright MCP alias map ships out of the
box, and server connections survive restarts via session-level restore. Ask the agent
$mcp
to list registered tags, or
$mcp-dump
to see the full tool map for the session.
Built-in skills like
$save,
$test, and
$browser
are just entries in a triggers array — add a trigger word and a BPI template to teach
the agent a new skill.
Dynamic tools passed per-request are hashed per session for cheap change detection. The same tool set is compiled differently per IDE — VS Code, Terax, and OpenCode each get their own grammar from one shared definition.
Pick a provider at startup. ZeroKey routes through the account you're already signed into — it never creates access, shares credentials, or stores your password. Switch providers anytime without touching your IDE config.
No separate DeepSeek API bill — your session persists across restarts, so you reconnect once and keep coding.
Use the Claude you already pay for in your IDE — cookies and org-scoped rate limits are handled for you.
Your ChatGPT plan, wired into your editor — sentinel tokens and conversation state are refreshed automatically.
Drop it into VS Code, Terax, or OpenCode in under two minutes. No plugins to pay for. You bring your own accounts.
http://localhost:7250/v1teraxhttp://localhost:7250/v1opencodeZeroKey is a self-hosted personal-use tool. You use your own accounts and credentials, and you're responsible for complying with each provider's Terms of Service.
No build step, no config files — one command:
pnpm install && pnpm start.
Pick a
Provider (DeepSeek, Claude, or ChatGPT), paste a
fetch()
copied from DevTools → Network tab as your User, then pick or create
a Session.
Set your IDE's base URL to
http://localhost:7250/v1
— full AI in your editor, no second bill.
# 1. clone, install, start >git clone https://github.com/downloaddoctor/zerokey.git >cd zerokey >pnpm install && pnpm start # 2. interactive wizard — pick provider, paste fetch, choose session # 3. call it like OpenAI >curl http://localhost:7250/v1/chat/completions \ -H 'Content-Type: application/json' \ -H 'Authorization: Bearer vscode' \ -d '{"messages":[{"role":"user","content":"hello"}],"model":"deepseek"}'
Objections handled up front, so you can decide in one read.
No. ZeroKey itself is free and open source. You still need your own DeepSeek, Claude, or ChatGPT account, but you don't pay for a separate API on top of it.
No. There's no API key involved. You paste a
fetch()
copied from your own browser's DevTools — ZeroKey only ever sees the session cookie
you choose to share, and it stays on your machine.
You're using your own account through your own browser session — but you are responsible for reading and complying with each provider's ToS. ZeroKey doesn't create access or bypass authentication.
Re-paste a fresh
fetch()
in the wizard — takes about 30 seconds. Sessions otherwise persist across server
restarts.
VS Code (built-in Chat), Terax, and OpenCode today — any OpenAI-compatible client
works against
localhost:7250/v1.
Yes. Restart the wizard and pick a different provider — no IDE reconfiguration needed.
Open source, self-hosted, no telemetry. If it doesn't work in 2 minutes, open an issue — it's free either way.