Free & open source · the handoff layer for AI work
Stop being the hallway.
Claude drafts. Codex edits. ChatGPT reviews. A teammate owns the call. The weak point is never the model — it's the handoff. Agent Relay turns every handoff into a visible task record that carries the source, the limits, the status, and the receipt.
# install the kit from GitHub, then start a relay
$ git clone https://github.com/The-Little-AI-Company/open-work-relay
$ cd open-work-relay
$ npm install -g .
$ relay init "Client Ops Relay"
Created relay workspace: Client Ops Relay
Next: relay dashboardPrompt mode vs work mode
A prompt asks for an answer. Work mode gives a job that survives the handoff.
Prompt mode is fine until the answer has to move to another person, agent, or system. Work mode carries everything the next actor needs.
# prompt mode
Write a follow-up email.
# work mode
Use the client-call transcript and the decision summary. Draft the follow-up, do not overstate the promise, flag what needs my judgment, leave a receipt, and stop before sending.
The task record
Seven parts travel with every job.
The next person or agent should resume from the record — not from a giant chat history. These seven parts are the whole contract.
- Outcome
- Owner / next actor
- Source material
- Context / decisions so far
- Allowed actions
- Stop rules / human gates
- Done evidence / receipt
The state machine
Work has a status anyone can read. Every change leaves a receipt.
Agents claim work, move it through statuses, and route anything ambiguous to needs-input. Receipts beat vibes.
The method, in our own framing
Three guides take you from idea to running loop.
The Relay Loop Audit
Nine questions that decide whether a handoff can leave the chat. Any “no” is the gap to fix before you hand off.
Read →02Prompt Craft
The climb from a prompt to a handoff-ready task: Ask → Frame → Bound → Hand off. One example, every rung.
Read →03The 30-Minute Relay
Six five-minute blocks from zero to a running loop on one real task. If it doesn’t fit in 30 minutes, the task was too big.
Read →