freeq ties AI model spend to identity, not shared API keys

Multi-agent systems face a mundane but thorny problem: when several AI agents need to spend model capacity on a human's behalf, sharing an API key is usually the only lever a provider offers — and that's both risky and unaccountable. freeq, a cryptographically-identified IRC-like protocol, closes that gap by treating spending authority as a delegatable, bounded, revocable object tied to identity rather than a secret.

The protocol already had signed budgets (per-channel spend limits with a named sponsor) and signed identity (DIDs for humans, certified keypairs for agents, delegation chains that narrow permissions as they pass down). What was missing was enforcement: the budget system carefully metered reported spend, but never touched the actual paid model calls, which ran through an unmediated proxy holding the provider's key.

This update joins the two systems. Model calls now go through an OpenAI-compatible endpoint that resolves the caller's delegation chain, checks the channel budget before making the upstream request, and returns HTTP 402 if the budget is exhausted — with no request ever reaching the provider. Charges are computed from the provider's real token counts rather than client claims, and unpriced models fall back to a deliberately expensive default to prevent free-riding.

For engineers building multi-agent systems, this is a concrete pattern for capacity delegation: revoking access means editing a budget instead of rotating a key that breaks every other consumer relying on it. Gaps remain, notably that sponsor consent still isn't verified — anyone can name any DID as the payer of their budget.