loom · Go · live demo
Agents that work together, on a real runtime.
loom is the orchestration layer I built under an AI storytelling product: versioned agents, multi-agent turns, hooks, session forks, budgets and a cost ledger. Host a multiplayer tabletop game: write the world, share a code, and let an AI game master run it, or run it yourself with AI-voiced NPCs. Every round shows what each agent cost. Or rewire an agent pipeline and watch loom schedule it.
When you host a game, your key stays in the demo server's memory for that game only, is never written to disk or logs, and pays for every AI step at your table. Players never need a key. Use a key with a low spend limit and revoke it afterwards. Each game is capped at 400 agent steps a day by loom's own budget guard. No key? Everything runs as a dry run with canned model output.
The world
Call for a check
Adjust a character
Make an NPC speak
You never declare dependencies. Each agent's template references other agents' outputs as {{.Inputs.Name}}; loom reads those references, checks every one has a producer, and plans the graph into layers that run concurrently. The first agent is the lead. Uses the model and key above.
Check the wiring to see how loom will schedule these agents.
- Run the pipeline to see each agent's output, tokens and cost.