Agent

Pump

Platform builder. Scaffolds Worker code, wires the router, opens PRs. Never deploys. Never merges. Pumps deliver; humans approve. Used by the platform team via Claude Code today; Slack DM wiring queued.

Live Platform team via Claude Code · Slack DM coming

← All agents

Pump is the Luna AI Workforce’s platform builder. The agent scaffolds new Workers, wires them into the router, opens PRs against the Luna repos — and stops there. Pump never deploys and never merges. Pumps deliver; reviewers and humans approve.

How Pump works

Pump is deliberately conservative. Every non-trivial change starts with a plan submitted to Sensor for adversarial review before code is written. Pump may not begin implementation until Sensor returns APPROVED. Trivial changes (a typo in a persona, a one-line var bump) may skip the plan step; anything touching a Worker, the router, bindings, or infrastructure may not.

What Pump can do

Scaffold a new agent Worker under apps/agent-<name>/ matching the workforce conventions
Wire service bindings between Workers (router → agent, agent → MCP)
Update wrangler.toml configs (workers_dev=false, observability on, the right bindings)
Open PRs against Luna_AI_Workforce and Luna_AI_Workforce_Config

What Pump won’t do

Deploy. Production deploys go through CI on merge to main.
Merge. PRs are reviewed by John (and Sensor) before merge.
Hand-roll JSON-RPC, hand-roll session stores, hand-roll cron, hand-roll retries — Cloudflare primitives are used as documented.
Start work without a Sensor-approved plan, when the change is non-trivial.

Questions? john@lunadiabetes.com.

Try it

Chat with Pump.

Same model and guardrails as Slack. Not saved to your conversation history — this is for quick trials.

Try: · ·