Integrations
MCP
Expose the control plane as tools an agent can call: mint a key, read usage, set a cap.
The MCP server wraps the control plane, not the gateway. An agent uses it to *administer*, provision a customer, check what someone spent, tighten a ceiling. Inference still goes through the base URL as normal.
Keep the two apart in your head: MCP is the console as tools; the base URL is the metered path. An agent that does inference over MCP is doing it wrong and will not be metered per end user.
Connect
bash
Tools
| Tool | What it does | Writes? |
|---|---|---|
keys_create | Mint an ak_ for one end user or agent | yes |
keys_list | Every key in the org, with actor label and cap | no |
keys_rotate | New secret for an existing key, old one revoked | yes |
usage_query | Spend by key, model or day | no |
cap_set | Change a key's ceiling and period | yes |
credits_grant | Add balance to a wallet | yes |
bundles_list | Bundles, allowed models and markup | no |
A session
text
Writing tools are refused unless the sk_ has the scope for them. A read-only key makes the whole server safe to hand to an agent you do not fully trust.