Relay

An orchestration engine, built from scratch

Give every agent
a place in the plan.

Turn a goal into connected steps. Run independent work in parallel. Keep every result, even when a worker stops.

No key needed for the demo. Bring your own key for real model runs.

Dependency schedulerIllustration
schedulingdurabilityagent_loopssynthesisWaiting for 3 dependencies

Independent branches run together. Each completed result is saved before dependent work can use it.

Dependency-aware schedulingPostgres checkpointsWorker crash recoveryPer-step traces

A workflow you can actually follow.

See the inputs, decisions and failures behind the final answer.

Inside the engine

One goal. A real graph.

A planner proposes the work. The engine validates dependencies and starts each step as soon as its inputs are ready.

Failures have a next step.

Retry temporary errors, repair a failed branch, or resume from the last saved result after a worker disappears.

Nothing hidden in the answer.

Inspect resolved prompts, attempts, sources, tokens and estimated costs. Export the report and execution trace.

Start with a working pattern.

Six editable workflows, from your first parallel run to a failure-recovery experiment.

Open playground

The engine coordinates.
The agents do the work.

Pi, Claude Code and Codex each bring their own model and tool loop. Relay provides the dependencies, checkpoints and failure policy around them.

Explore the local adapters

Pi

JSON event stream

Read-only analysis through Pi’s local CLI.

Claude Code

Print JSON

Programmatic agent results and reported usage.

Codex

Exec JSONL

Ephemeral, read-only runs with structured events.

Local adapters run on your machine. The hosted playground uses model APIs.

Your workflow. Your provider.

Connect OpenAI, Anthropic, Gemini, Groq, Cerebras or OpenRouter. Keys are encrypted, and your browser workspace stays private.

Connect a provider