How auto-atlas/atlas works
Atlas is presented as a private, local-first assistant whose voice loop is the front door: a user speaks, Atlas listens and responds, and it can gate tools or delegate specialist work and notify the user when it finishes. The repository also contains a substantial OpenJarvis runtime and frontend; the supplied excerpts show OpenJarvis ingestion, telemetry, workflows, and engine dispatch, but not their integration with the Atlas voice loop. README statements support the Atlas workflow links below; broader internal wiring is left out rather than inferred.
Main entry points: voice_llm.py Voice loop · agent_bridge.py Agent hub · morning_digest.py Morning brief · StreamService.kt Phone and watch
Components 20
Voice assistant
- Voice loop
voice_llm.py - Voice transport
app/src/openjarvis/agents - Local inference
app/src/openjarvis/engine - Trust tiers
app/src/openjarvis/security
Tools and agents
- Assistant tools
app/src/openjarvis/tools/_stubs.py - Agent hub
agent_bridge.py - Agent fabric
a2a_fabric.py - Agent tasks
agent_tasks.py - Agent results
agent_result_store.py - Async delivery
agent_delivery.py
Memory and integrations
- Persistent memory
app/src/openjarvis/connectors/store.py - Data connectors
app/src/openjarvis/connectors/pipeline.py - Morning brief
app/src/openjarvis/agents/morning_digest.py
User surfaces
- Web interface
app/frontend/src/App.tsx - Phone and watch
eve-app/android/app/src/main/java/app/eve/push/StreamService.kt - Glasses bridge
glasses/mentra-eve/src/eve-link.ts
Other components
- User
- Obsidian wiki
- Specialist agents
- Local hardware
Connections 17
- User → Voice loop: speaks to
- Voice loop → Local hardware: runs locally on
- Voice loop → Trust tiers: gates risky tools
- Voice loop → Assistant tools: uses local tools
- Voice loop → Agent hub: hands off work
- Agent hub → Specialist agents: delegates work
- Agent hub → Agent tasks: tracks delegation
- Agent tasks → Agent results: stores results
- Agent hub → Async delivery: notifies on completion
- Voice loop → Persistent memory: reloads session facts
- Data connectors → Persistent memory: ingests documents
app/src/openjarvis/connectors/pipeline.py - Morning brief → Obsidian wiki: grounds strategy
- Morning brief → Specialist agents: delegates strategy
- Phone and watch → Voice loop: streams voice
- Phone and watch → Async delivery: receives approvals
- Glasses bridge → Voice loop: bridges camera and audio
- Web interface → Voice loop: offers assistant UI