cerber

A trust-first, multi-provider AI gateway. Speaks the OpenAI and Anthropic API dialects, pools multiple accounts per provider, and routes each request to the right upstream.

Contents

Authentication

• From the local network (loopback/LAN): typically no key required.
• From outside: send Authorization: Bearer <KEY> or x-api-key: <KEY>.
• Keys are minted & managed by the operator in the dashboard. Invalid/missing key → 401.

API endpoints

Embeddings/completions/responses are served only by OpenAI-compatible providers (OpenAI/Grok/ollama). A model that routes to Anthropic, or an unknown model, → 400; a provider lacking the capability → 501.

Two dialects, one gateway

Point an OpenAI SDK at https://cerber.ihatebot.com/v1 or an Anthropic SDK at https://cerber.ihatebot.com. cerber translates OpenAI⇄Anthropic and OpenAI⇄Gemini under the hood, so you can call a Claude model through the OpenAI endpoint and get OpenAI-shaped responses, or use the native Anthropic endpoint for byte-faithful passthrough.

Model routing & aliases

Just set model; cerber routes by name:

Configured routing prefixes and discovered model names take precedence; an unmatched model → 400.

Automatic fallback (OpenAI endpoint)

On /v1/chat/completions, if the model fails with a retryable error — provider 5xx, or every account rate-limited/unavailable — cerber retries the next model in a fallback chain before any bytes are sent. A 4xx client error is returned as-is, and a stream already in progress is never re-attempted.

Per-request override: header X-Cerber-Fallback: model2,model3 (each routed like any model, so it can cross providers).

No server-side chains configured on this instance; use the header.

Rate limits & budgets

A managed key may carry a rolling cost budget (USD, from configured model pricing) and rate limits (requests & tokens per window). On exhaustion you get:

Windows are rolling: minute / hour / day / week / month. Counters persist across restarts. Streamed responses currently count requests but not cost/tokens.

Request headers

Providers & models on this instance

anthropic (3 models)

arliai (59 models)

gemini (39 models)

grok (10 models)

ollama

Use this provider's standard model names (routed by the prefixes above).

openai

Use this provider's standard model names (routed by the prefixes above).

Admin API

⚠ No management key configured — /admin/* currently accepts any valid client key. Set access.management_key for multi-tenant deployments, otherwise a capped key could lift its own limits.

Observability

• GET /metrics — Prometheus metrics (counts + credential names, no secrets).
• GET /dashboard — usage, accounts, total cost, per-hour chart, recent-requests table, key management.
• GET /admin/requests — recent per-request log (who/IP/UA/model/tokens/cost), in-memory ring (not persisted, holds IPs).
• GET /admin/stats & /admin/usage.csv — JSON/CSV snapshots; total_cost is the cumulative $ of all priced tokens through cerber.

Examples

OpenAI dialect (curl)

curl https://cerber.ihatebot.com/v1/chat/completions \
  -H 'Authorization: Bearer $KEY' -H 'Content-Type: application/json' \
  -d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"hi"}]}'

With per-request fallback

curl https://cerber.ihatebot.com/v1/chat/completions \
  -H 'Authorization: Bearer $KEY' -H 'X-Cerber-Fallback: gpt-4o,gemini-2.0-flash' \
  -d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"hi"}]}'

Anthropic dialect (curl)

curl https://cerber.ihatebot.com/v1/messages \
  -H 'Authorization: Bearer $KEY' -H 'anthropic-version: 2023-06-01' -H 'Content-Type: application/json' \
  -d '{"model":"claude-sonnet-5","max_tokens":256,"messages":[{"role":"user","content":"hi"}]}'

OpenAI Python SDK

from openai import OpenAI
client = OpenAI(base_url="https://cerber.ihatebot.com/v1", api_key="$KEY")
client.chat.completions.create(model="claude-sonnet-5",
    messages=[{"role":"user","content":"hi"}])

Trust principles

• No phone-home. cerber makes outbound calls only to the provider APIs a request is routed to — no telemetry, no analytics, no remote asset fetch.
• Credentials never leave the box except as the documented auth header to the provider that owns them.
• Pinned, opt-in, never silent. No feature fetches and runs remote content automatically.

cerber · github.com/tggo/cerber