Agent Relay is a local dashboard and relay workspace for moving work between people and AI agents without losing the source, limits, status, or receipt.
It starts from a simple idea:
Prompt mode asks an AI for an answer. Work mode gives a job to a system that can carry state, stop, resume, and show evidence.
Most AI users are already working across several tools. Claude might draft. Codex might edit files. ChatGPT might review. A browser agent might verify. A teammate might own the final call. The weak point is not the model. The weak point is the handoff.
Agent Relay turns the handoff into a visible task record.
This repo is the starter kit and first local product:
- a setup prompt,
- an
AGENTS.mdprotocol block, - a formal specification,
- a Node CLI,
- a local dashboard,
- a task template,
- a workflow template,
- a receipt template,
- a Markdown-folder queue,
- adapter notes for GitHub Issues, Trello, Linear, and Notion,
- a smoke test,
- platform-specific launch drafts.
- Not another private assistant.
- Not a replacement for judgment.
- Not a required project-management tool.
- Not a hidden paid protocol.
- Not a clone of anyone else's paid guide.
Agent Relay is an independent product built around a public problem: humans should not be the copy-paste path between agents.
- Open the setup prompt:
prompts/agent-relay-setup-prompt.md - Paste it into the AI agent you want to configure.
- Answer the setup questions.
- Pick your primary queue:
- plain Markdown folder,
- GitHub Issues,
- Trello,
- Linear,
- Notion,
- or another reachable queue.
- Let the agent propose workflow families for your real work.
- Add the generated protocol to your repo's
AGENTS.md. - Run the smoke test:
examples/markdown-folder/tasks/0001-say-hello-from-the-relay.md
The first product surface is a local personal dashboard for the user's projects.
Create a named relay workspace, then open its dashboard — two short commands, no flags:
relay init "Client Ops Relay"
relay dashboardinit remembers the relay it just created, so dashboard finds it with no --root. You can also cd into any relay folder and run relay dashboard there.
Install the relay command globally from a clone:
git clone https://github.com/The-Little-AI-Company/open-work-relay
cd open-work-relay
npm install -g .Or run it without installing from inside the clone:
node bin/agent-relay.js init "Client Ops Relay"
node bin/agent-relay.js dashboardFrom a clone, the same commands are node bin/agent-relay.js init "..." and node bin/agent-relay.js dashboard.
The dashboard uses the relay's configured name, so the user sees "Client Ops Relay" or whatever they named it, not a hard-coded product label.
The dashboard uses the Ink Console direction: project visibility, work status, human attention, and receipts in one local browser view.
If no --root is provided during init, relay state still lives outside project repos by default:
~/.agent-relay/
client-ops-relay/
relay.json
tasks/
receipts/
done/
Use a different workspace when testing:
node bin/agent-relay.js dashboard --root ./scratch-relay --port 8788Prompt mode:
Write a follow-up email.
Work mode:
Use the client-call transcript and the decision summary. Draft the follow-up email, do not overstate the promise, flag anything that needs my judgment, leave a receipt, and stop before sending.
The second version is work. It has source material, boundaries, ownership, evidence, and a stop rule.
The relay should not make every job look identical, but the first examples should stay close to the public transcript.
The transcript-grounded workflow families are:
- client call / follow-up,
- support ticket,
- schedule / pickup change.
Examples:
examples/workflows/client-call-follow-up.mdexamples/workflows/support-escalation.mdexamples/workflows/household-schedule-change.md
Agent Relay extensions can add planning, production, branding/content, or custom workflows after the core is clear.
Extension examples:
The full method — the formal specification, product principles, the Relay Loop Audit, Prompt Craft, the 30-Minute Relay, and source grounding — lives in a private companion repository. The kit in this repo is everything you need to run a relay; the method writeups are available separately.
Every task uses the same core shape:
- Outcome
- Owner / next actor
- Source material
- Context / decisions so far
- Allowed actions
- Stop rules / human gates
- Done evidence / receipt
Optional metadata helps agents pick the right task:
- priority,
- deadline,
- project,
- tags,
- source queue,
- external issue URL or ID,
- dependencies,
- blocked reason.
See:
templates/task.md
Workflow template:
templates/workflow.md
Use these statuses unless your queue has native equivalents:
readyclaimedworkingneeds-inputreviewdoneparked
Every status change should leave a short receipt.
See:
templates/receipt.md
When an agent stops, it should say why.
## Stop
Status: needs-input | review | done | parked
Reason:
What I completed:
What I did not do:
What I need next:
Next actor:
Receipt:The stop reason matters because the next person or agent should be able to resume without guessing.
The protocol is queue-agnostic. The queue only has to be readable and writable by the human or agent doing the work.
Adapters:
adapters/markdown-folder.mdadapters/github-issues.mdadapters/trello.mdadapters/linear.mdadapters/notion.md
Markdown is the universal fallback because it is transparent, portable, and easy for any agent to inspect.
The core kit is free because useful AI knowledge should travel.
If this saves you a subscription or a few hours of copy-paste work, consider supporting Jeff's work. The URLs, servers, tools, and hours still cost real money.
Support link: TODO
This project was sparked by public discussion about AI agent handoffs and the real pain of moving work across Claude, Codex, ChatGPT, OpenClaw, Hermes, browser agents, and team tools.
No paid material was used to build this kit. The transcript-grounded patterns are documented in the private method repository. The queue adapters, templates, and TLAC packaging are independent Agent Relay implementation choices.
MIT. Use it, adapt it, teach it, remix it.