Connect your AI

The AI you already talk to can boot and run bootkey agents for you. Add bootkey to Claude, ChatGPT, Cursor, Claude Code or any other app that speaks MCP, sign in once, and then ask in plain words: “boot me an agent that watches my servers”, “how is Sentinel doing?”, “give Relay a duty every weekday at 8”.

Each agent it boots is a real one: its own Linux server with its own IP, web address, email address and memory, thinking on your own OpenRouter or Venice key. Your AI does the setup. The agent keeps running after the chat ends.

Add it

The address is:

https://bootkey.ai/mcp

Your app opens a bootkey page. You sign in (or make an account), see what the app will be able to do, and connect. That’s all.

Claude (claude.ai, desktop and mobile): add a custom connector in Settings, under Connectors, with the address above.
ChatGPT: turn on developer mode in the connector settings and add a connector with the address above and OAuth sign-in.
Claude Code: claude mcp add --transport http bootkey https://bootkey.ai/mcp, then /mcp to sign in.
Cursor, VS Code, Windsurf and others: add a remote MCP server with the address. In Cursor’s mcp.json:
```
{ "mcpServers": { "bootkey": { "url": "https://bootkey.ai/mcp" } } }
```

Before your AI can boot its first agent, the account needs a model key (OpenRouter is one click in the console), an SSH key and some balance. The connect page tells you which of these is missing, and your AI will too.

What it can do

It can	It can’t
Boot new agents, up to 3 a day, on your balance ($9 a month each, billed by the hour)	Destroy, hibernate or wake an agent
Set up your agents: duties, objectives, the files that shape its job, its vault, temper, model and the machines it may log into	Touch your card, your model keys, your team or your keys
Talk to your agents and read their replies	Connect other apps or make setup keys
Add money in USDC from a wallet it controls, when you ask it to	Do anything you can’t see

When you connect, you choose whether it looks after all your agents or only the ones it boots. Every change it makes shows up in that agent’s activity log, for example: Claude added the duty “orders”, through its bootkey connection. You get an email when an app is connected, and when one adds money.

Tools

Tool	What it does
`get_account`	Balance, model keys, SSH keys, every agent, and what this connection may do. Your AI starts here.
`get_options`	The jobs an agent can start with, and the models with prices and leaderboard ranks.
`read_manual`	The full manual or the API reference.
`boot_agent`	Boots an agent. It comes up in about two minutes.
`get_agent`	Status, web address, email, what it’s doing now, its latest line, vault names, recent activity.
`talk_to_agent`, `get_reply`	A message to the agent, and its answer. Long tasks can take minutes.
`agent_command`	Setup on the box: files, duties, model, machines, logs.
`set_temper`	Tool, worker, worker by day, or alive.
`save_secrets`	Keys and tokens for its vault. They can’t be read back.
`pay_with_usdc`, `check_payment`	An invoice your AI pays from its own wallet, and whether it arrived.

Paying in USDC

If you want your AI to pay for its agents, it can. pay_with_usdc opens an invoice for $10 to $500 on Base, Polygon, Solana or Ethereum, with a deposit address used for that payment only and the exact amount to send. Once the network can’t undo the transfer (about a minute on Solana, up to 20 on Base), the money is on your balance and you get an email saying who sent it. Only Circle’s native USDC counts.

Opening an invoice costs nothing, and your AI can’t spend anything from your account except by booting agents, which it does only up to 3 a day.

Without a connector

Any AI that can run curl can do the same with a setup key: Copy agent prompt in the console makes a key that lasts an hour and puts it in a prompt you paste into your AI. See the API.

Turning it off

Disconnect an app on your Security page, under Connected apps. Its access ends at once. Removing the connector in the app ends it too.

For developers

bootkey’s MCP server speaks Streamable HTTP and answers with plain JSON (no sessions, no event stream). Sign-in is OAuth 2.1:

Discovery: /.well-known/oauth-protected-resource/mcp (RFC 9728) and /.well-known/oauth-authorization-server (RFC 8414). A request without a token gets a 401 whose WWW-Authenticate header points there.
Registration: POST /oauth/register (RFC 7591). Public clients use token_endpoint_auth_method: "none".
Authorization code with PKCE (S256 only) at /oauth/authorize, tokens at /oauth/token, revocation at /oauth/revoke. Loopback redirects may use any port.
Access tokens last an hour. Refresh tokens last 30 days and change on every use; an old one used again ends the connection.
The one scope is agents. The resource is https://bootkey.ai/mcp.

A setup key (Authorization: Bearer bk_key_…) also works on /mcp, for an hour.