Jarvis
A voice-first chief of staff for a one-person studio. The architecture is public.
Voice-first AI chief of staff that runs the agent fleet. Architecture is public.
Status: In development · studio · built and run by Jack Stovell at Adapt Progress Evolve
What it is
Jarvis is the voice-first AI assistant that runs my studio, backed by a fleet of Claude agents for memory, research, security, content, design review and more. The system itself is private because it runs my life. The architecture, the numbers and the failures are all public on GitHub.
Who it's for
Built for me, and published for developers who want to see how a personal multi-agent system gets put together, including what broke.
How it works
- Talk to it. Speak an ask, and Jarvis either answers or hands the work on.
- Agents do the work. Specialised agents coordinate over a shared message bus with persistent memory.
- Memory as a graph. What Jarvis knows lives in a knowledge graph, visualised with Cortex Map.
- Human approval. Anything that ships waits for a yes from me.
Timeline
- — Cortex Map released from Jarvis's memory view
- — Architecture published on GitHub
Facts
- Voice latency (p50)
- 3.0s over 847 turns
- Status
- In development
- Architecture
- Public on GitHub
Under the hood
Stack: Node, Postgres, Claude, Voice.
Agents in the fleet that work on it: Jarvis (voice), Memory, Planner, Infra, Docs, Calendar, Morning brief, Context, Reconciler, Voice QA, Dashboard, Coordinator.
FAQ
What is Jarvis by Jack Stovell?
Jarvis is my personal, voice-first AI assistant, and it coordinates a fleet of Claude agents to run Adapt Progress Evolve. Its architecture is documented publicly at github.com/StovBuilds/jarvis-architecture.
Can I use Jarvis?
No, Jarvis isn't available to use. It's single-user and private. The architecture reference is public so other people can learn from the design.
How fast does Jarvis respond by voice?
The public Jarvis architecture reference measured a median voice response of 3.0 seconds and a 90th percentile of 6.3 seconds over 847 real conversational turns.