As your agents multiply, they need their own identity, their own credentials, their own memory, their own audit trail — isolated from you and from each other. Today, they borrow yours. That doesn't scale, and it doesn't survive the first incident.
Install, sign in, store a key. Your agent reads via MCP. If anything goes wrong, one command kills access across every service that agent was touching — in under six seconds.
The skill route gives the agent agency — it asks for credentials when it needs them, provisions what's missing, and releases them when it's done. The launch route is lower-magic: your settings.json stays untouched, and AgentKeys fills in env vars at launch.
Bring your own Claude or OpenRouter key. From there, your OpenClaw agent can ask AgentKeys for anything else it needs — OpenAI TTS, Brave Search, SendGrid — and we'll sign it up, verify the email, and hand back a scoped key. No human, ~30 seconds per service.
We hand your agent a key and get out of the way. Your agent calls OpenRouter, GitHub, Notion directly — we don't proxy traffic, so your latency and reliability stay in your hands.
At launch you'll pay in dollars, yuan, or euros — AgentKeys handles the currency conversion behind the scenes to pay service providers. Your agent sees a balance; you never see a wallet. During internal test, we're running on test credits.
AgentKeys is MIT OR Apache-2.0. Every component in the trust boundary is open source — CLI, daemon, MCP adapter, provisioner, the credential backend itself.
Reproducible builds verified in CI. Supply chain scanned with cargo audit and cargo vet.
We're honest about limits: on a stock sandbox, the session file is readable by the sandbox's own user — so fast revocation (under six seconds) is our primary line of defense. We don't claim SOC 2 we don't have.
AgentKeys is in internal test. Request early access and store your first credential in under a minute.