Jarvis

The suite

Jarvis

About this suite

Jarvis is a private AI operating layer: one front door, two surfaces. Same kernel, same policy, two windows — the CAO conversation in Comrade, the read-only fleet observation in Fleet. Switch between them in the header.

Comrade

The conversation surface

Direct line to the CAO. Send a message, ask a question, get an answer back. The model picker is on /; conversation history lives in the left rail; Ask-First items land inline as cards you can answer without leaving the thread.

Fleet

The observation surface

Mirror of the running fleet: uptime, subagents, event log, cron, approvals, channels, continuity, recall. Read-only by design — Glass never grants merge or mutation authority. Intervention rail on Runs and Cron is policy-gated, audit-logged, and routes through harnessd.

How Jarvis is built

  • Policy before action. Every side effect goes through a capability grant. No ambient authority. The user defines what the agent can touch; the kernel enforces it.
  • Same front door, two surfaces. Comrade is for the operator. Fleet is for the observer. Both read the same kernel state; neither of them is the authority.
  • Memory under your control. The brain is local-first. Indexing and recall are queryable; nothing leaves the box without an explicit outbound.
  • Multi-model by default. Different models for different agents, different tasks, different cost/latency/sensitivity profiles. Routing is policy, not habit.
  • Native SOBER + PRISM. Doctrine-aware code review (SOBER) and a seven-tool sovereign harness (PRISM) ship with the project.

Where to go next