Relay
A feedback engine that reads how you work, finds what you repeat, builds skills and MCP connectors from it, and iteratively improves them with your approval.
What it is
A feedback engine that watches how you work, finds repeated patterns, generates skills and MCP connectors, and iteratively improves them with your approval so the harness gets better over time.
Why it is here
It is the piece that closes the loop between capturing how work happens and actually making the system better from it. It is running internally and it is incredibly useful.
What it is
Relay captures local activity data — keystrokes, app usage, behavioral patterns, everything happening on the machine. It takes the skills you have already built around that data, looks at what you are doing, and suggests new skills and ways to extend existing ones to take more work off your plate.
The other half of the stack is the data scraping layer and the workhorse API — that part captures how work actually happens and turns it into structured telemetry. Together, they close the loop.
What it does
- Orchestration. Live tree of every agent, sub-agent, and cloud worker. What is working, what is idle, what is blocked.
- Spend. AI cost per agent, per model, per workstream. OpenRouter, Anthropic, whatever you are using.
- Cloud. How much remote compute is live and what it is doing.
- Team. See what other operators' agents are doing in the same tree.
- Skill suggestions. The telemetry cycles back around and generates suggestions for new skills and MCP connectors based on what you actually repeat. It does not create the skill directly — it creates a suggestion. You review it, approve it, edit it if you want.
- Iterative updates. As a manager, you can keep checking and updating those skills. The system continues to suggest ways to add to them so the agent does more and the person does less over time.
Live agents, blocked work, automations, AI spend, cloud capacity.
One operator's tree — agents, sub-agents, apps, and live state.
Multiple operators and their agents in one view.
Discovery queue, skill packaging, deployment path.
How it connects
The data engine captures work — app usage, keystrokes, context switches, workflow patterns — and generates reports, briefs, and suggestions.
Relay takes those suggestions and turns them into deployable skills, MCP connectors, AGENTS.md behaviors, and automations. Then it tracks whether they helped and feeds that back into the next cycle.
The skill-builder repo is the open-source standalone version of the packaging side. Relay folds that into the full system.
Skills are markdown files and MCP servers on open standards — they run across Claude, Codex, GPT, Gemini.
Where it is
Running internally across multiple operators and agents every day.