One engine, type-checked at config-load.

Available variables

Variable	Type	Notes
request	Request	model, messages, headers, backend, path, size, stream
tenant	Tenant	id, plan, region, flags
user	User	id, email, role (from auth header)
quota	Quota	remaining, limit, resets_at
receipt	Receipt	id, tokens_in, tokens_out, cost, annotations
time	Timestamp	current time (seconds since epoch)

Built-in functions

• allow(), deny(reason): pre-request disposition.
• annotate(key, value): attach metadata to the receipt.
• route(upstream_id): pick the upstream for this request.
• rate_limit(key, per_minute, per_hour): enforce a rate cap, keyed arbitrarily.
• quota_spend(amount): deduct from the quota ledger.
• redact_field(path, pattern): mutate a specific field.
• challenge_402(price, currency): issue an x402 payment challenge.

Operator semantics

Standard CEL rules apply: short-circuit && and ||, ternary ?:, membership with in, string methods like startsWith and contains. No loops. No I/O. Expressions are pure; side effects come from the built-in functions above.

Type checking at load

When you run relaygate config apply, every policy is compiled and type-checked. Type errors fail the load. The error message points at the file, line, and token:

error: type mismatch in policy 'route-by-plan'
  /etc/relaygate/policies/route.cel:4:18
  tenant.plan == 'pro' && request.modl == 'gpt-5'
                                 ^^^^
  no such field 'modl' on type Request
  did you mean 'model'?

Examples

1. Simple deny

tenant.plan == "free" && request.model == "gpt-5"
  ? deny("gpt-5 is not on the free plan")
  : allow()

2. Route by tenant plan

tenant.plan == "pro" ? route("anthropic-direct") : route("openai")

3. Per-tenant rate-limit

rate_limit(
  key = "tenant:" + tenant.id,
  per_minute = tenant.plan == "pro" ? 1200 : 60
)

4. Quota enforcement

quota.remaining <= 0
  ? deny("quota exhausted; resets at " + string(quota.resets_at))
  : quota_spend(1)

5. Content-based routing

request.messages.size() > 0
  && request.messages[request.messages.size() - 1].content.contains("code:")
  ? route("anthropic-direct")
  : route("openai")

6. Time-window policy

// between 22:00 and 06:00 UTC, route to the cheaper backend
time.getHours() >= 22 || time.getHours() < 6
  ? route("openrouter")
  : route("openai")

7. Role-based redaction

user.role != "admin"
  ? redact_field("request.messages[*].content", "\\b\\d{3}-\\d{2}-\\d{4}\\b")
  : allow()

8. x402 commerce challenge

request.path.startsWith("/v1/agents/")
  && !request.headers.exists("x-truecom-settlement")
  ? challenge_402(0.004, "USDC")
  : allow()

Next: Observability →