---
title: Hep.gg AI (LLM)
description: An Anthropic-compatible Claude proxy, an OpenAI-compatible chat endpoint for self-service LLM keys, and a read-only usage API.
---

# Hep.gg AI

Two programmatic APIs share one account: an Anthropic-compatible proxy that
meters Claude usage per machine, and an OpenAI-compatible chat endpoint for
self-service LLM keys. A read-only usage API exposes cached account usage.

Base URL: https://ai.hep.gg   (every path is also mirrored under https://hep.gg/ai)

## Anthropic proxy (Claude)

### POST /anthropic/:proxy_token/v1/messages
Anthropic Messages API, forwarded transparently. Use the proxy token as the URL
segment (and optionally as Authorization: Bearer <proxy_token>). SSE streaming
works when the body sets "stream": true. Most Anthropic headers (model, beta
features) pass through; sensitive headers are stripped. Other Anthropic paths
under /anthropic/:proxy_token/... (for example /v1/messages/count_tokens) are
forwarded the same way. Proxy tokens are issued per machine and managed by the
ccas CLI / the dashboard.

Errors use the Anthropic shape, e.g. on an unknown token:
  { "type": "error", "error": { "type": "authentication_error", "message": "..." } }

## OpenAI-compatible API (self-service LLM keys)

Auth: Authorization: Bearer sk-hyd-<...>   (mint sk-hyd keys on the dashboard
LLM Keys page). Non-admins may call user-allowed models (e.g. qwen3-8b).

### POST /v1/chat/completions
OpenAI Chat Completions shape: body { model, messages, stream?, ... }. SSE
streaming when "stream": true. Past the rolling 30-day USD cap this returns
429 with an OpenAI-style error body:
  { "error": { "message": "...", "type": "quota_exceeded", "code": "<reason>" } }
where reason is one of quota_exhausted | overage_cap | no_hep_tokens. Opt into
metered overage on the dashboard to keep serving past the cap.

### GET /v1/models
List the models your key may call. A key is bound to its model(s) at creation,
so this typically returns the single model that key is allowed to use.

## Usage API (read-only)

Auth: a per-user usage token, usg_<...> (mint with "ccas usage-token" or on the
dashboard). It only reads cached usage and can switch the active pooled account.

### GET /usage
Cached usage for every pooled account: { version, fetched_at, accounts: [ ...,
each carrying an "active" flag ] }. Supports ETag / 304.

### GET /usage/:id
One account's snapshot by stable id.

### POST /usage/accounts/:id/activate
Switch the globally-active pooled account.

## AI Skills (skills.hep.gg)

A separate registry of Claude-compatible agent skills served as raw markdown at
https://skills.hep.gg: fetch public skills, reach your private skills via a
capability URL, or save a skill with POST /me/:accessKey/import. Full reference:
https://docs.hep.gg/llms/ai/skills

## Hep.gg dashboard (cookie auth - not for programmatic use)
  GET    /api/v1/ai/account
  POST   /api/v1/ai/account/rotate
  GET    /api/v1/ai/usage-token        POST /api/v1/ai/usage-token   (reveal / rotate)
  GET    /api/v1/ai/machines           (machines are created by the ccas CLI)
  GET    /api/v1/ai/llm-keys
  POST   /api/v1/ai/llm-keys           { label?, model? }
  POST   /api/v1/ai/llm-keys/:id/rotate
  DELETE /api/v1/ai/llm-keys/:id
  GET    /api/v1/ai/accounts           (pooled ccas accounts + live usage)