Relay is the orchestration console for AI agents in production. Route, observe, and recover every run from one surface — without rewriting your loop.
No credit card required · Deploy on your own infra
Four primitives cover the full lifecycle of a production agent run. Compose them, don't bolt them on.
Route every request by capability, cost, and latency. Cache hits skip the expensive path automatically — your bill drops without touching your code.
Every tool call, token, and decision is a first-class trace. Drill from a failed run to the exact reasoning step that caused it.
Timeouts retry with backoff, failed tools degrade to fallbacks, and interrupted sessions resume from checkpoint — not from scratch.
Per-session and per-day budgets enforced at the request layer. When the budget is gone, the run stops — never silently overruns.
DeepSeek, Anthropic, OpenAI, and your own gateway — one interface, one telemetry stream. Bring your keys, keep your data.
Approvals, pauses, and handoffs where they matter. Agents run autonomously inside the envelope you define.
Relay wraps your existing agent — DSH, Claude Code, or your own loop — behind a thin config layer.
# relay.yaml — start with defaults, tune as you grow session: budget: $5.00 # hard cap, enforced pre-request resume: true # checkpoint on interrupt routing: default: deepseek-v4 # cost-optimal path fallback: [ deepseek-v3, gpt-oss ] cache: read # reuse prior turns tools: timeout: 30s retry: 2 degrade: true # fall back on partial failure telemetry: traces: true export: ./relay-runs.jsonl
Deploy in minutes, tune in days, trust from the first run.
Get started free