Open-source agent runtime

The operating layer
for autonomous
AI agents.

Memory. Awareness. Action. Orchestration. — Four primitives that turn any LLM into a persistent, context-aware agent that sees your screen, controls your apps, and pursues goals on your behalf. Provider-agnostic. Self-hosted. Runs 24/7.

Get Started View Source
$ bun install -g @usejarvis/brain
Discord Community Read the Docs View Roadmap
Runtime Status
--:--:-- UTC
Memory
vault 847 facts 12 entities · semantic search
context injected per-message · knowledge graph
Awareness
capture active ocr + vision · 7s cycle
context tracking activity sessions · struggle detection
Action
browser ready chromium 121 · cdp:9222
sidecar 3 online go · jwt-ws · multi-machine
voice listening wake-word · tts:edge
Orchestration
agents 11 roles 2 active · parallel delegation
workflows 4 active 2 cron · 1 webhook · 1 screen
goals 3 active 1 objective · 2 key results

The Four Primitives

Memory
Persistent semantic memory
Entity graph, knowledge vault, fact extraction after every conversation. Context continuity across sessions. Semantic search injection per message.
Awareness
User context understanding
Screen capture + OCR every 7s. Application state tracking, activity sessions, struggle detection. Cloud vision escalation for complex contexts.
Action
Execution across desktop, browser & tools
CDP browser automation. Go sidecar for native desktop apps. Terminal, filesystem, voice (TTS/STT). 14+ builtin tools with 200 iterations per turn.
Orchestration
Multi-agent coordination & autonomy
11 specialist agent roles with delegation. Visual workflow builder (50+ nodes). OKR goal pursuit. Authority gating, approval logic, proactive routines.
ARCHITECTURE

The assistant is the first
application. Not the last.

Most AI tools ship an assistant and call it a product. JARVIS ships a runtime — and the personal assistant is just the first application built on it. The same four primitives power every workflow, every agent role, and every capability on the platform.

Applications
Assistant built-in
Workflows built-in
Agent Roles 11 roles
Plugins SDK
Your App build it
Runtime
Memory
Awareness
Action
Orchestration

The runtime provides the layer that agents need to be persistent, context-aware, and capable of acting in the real world. Applications are built on top — the assistant is just the most visible one.

ACTION

One brain.
Every machine.

Install the sidecar on any computer — work laptop, home desktop, cloud VM, CI server. Each connects back to a single JARVIS daemon via JWT-authenticated WebSocket. One agent runtime orchestrating across all your machines.

Work Laptop terminal, browser, desktop
CI Server terminal, filesystem
JARVIS
DAEMON
Home Desktop desktop, filesystem, screenshot
Cloud VM terminal, browser, filesystem
# Connect any machine in 30 seconds
$ bun install -g @usejarvis/sidecar
$ jarvis-sidecar --token eyJhbG...

✓ Connected to jarvis daemon (ws://brain:3142)
✓ Capabilities: terminal, browser, desktop, filesystem
✓ Machine "work-laptop" online
# Orchestrate across machines
"Deploy the build on the CI server,
then open the staging URL on my laptop
and take a screenshot for the PR."

ci-server ✓ build deployed · 2m 14s
work-laptop ✓ browser opened · screenshot saved
daemon ✓ PR comment posted with image
∞ Machines connect any number of computers, VMs, or servers
0 Capabilities terminal, browser, desktop, filesystem, clipboard, screenshot, awareness
0 Brain single daemon orchestrates all connected sidecars
0 Config JWT token is all you need — sidecar auto-discovers capabilities
Sidecar Docs View Source
ORCHESTRATION

Agents that delegate
to agents.

The primary agent doesn't do everything itself. It delegates to 11 specialist roles — each with its own tools, context, and execution loop. Research runs in parallel with coding. Analysis happens while writing. Authority stays with the primary agent.

Primary Agent authority
Researcher web_search, browser, extract
Coder terminal, file_io, git
Writer file_io, template, format
Analyst data, calc, visualize
+ 7 more: sysadmin, designer, planner, reviewer, data-engineer, devops, security
# Delegate complex work to specialists
"Research competitor pricing, then write a report"

delegate researcher · working
tools: web_search, browser, extract
delegate writer · waiting
depends_on: researcher
# Parallel execution with dependency resolution
researcher completed 3 sources · 12 data points
writer drafting context injected from researcher
writer completed report → ~/reports/competitive.md

# Authority enforcement
writer denied send_email (governed action)
primary approved send_email → team@company.com
0 Specialist Roles researcher, coder, writer, analyst, sysadmin, designer, planner, reviewer, and more
0 Iterations / Turn agents run until done, not until the response looks done
∞ Parallel multiple specialists work simultaneously on different sub-tasks
0 Authority sub-agents inherit memory and awareness but cannot approve governed actions
Explore Agents Authority Docs
ORCHESTRATION

Orchestration,
not scripting.

Describe a workflow in plain English — or build it visually. 50+ nodes. Every trigger type. Self-healing execution. Workflows are first-class citizens of the runtime — they share memory, awareness, and action with every other agent.

Workflow Canvas
Cron · 9am daily
Agent: Summarize PRs
If urgent
Telegram alert
Slack summary
# Create via chat — just describe it
"Every morning at 9, check GitHub PRs.
If anything is urgent, Telegram me.
Otherwise, post a summary to Slack."

Workflow created · 5 nodes · 4 edges
Triggers active · cron: 0 9 * * *
# Self-healing execution
step 1 passed trigger.cron
step 2 passed action.agent_task
step 3 failed action.http_request
self-heal → fixed auth header
step 3 passed action.http_request (retry)
step 4 passed action.notification
0 Triggers cron, webhook, file, screen, git, email, calendar, clipboard, process, poll, manual
0 Actions HTTP, agent task, shell, code, Gmail, Telegram, Discord, notification, file write, run tool
0 Logic if/else, switch, loop, delay, merge, race, variable set/get, template render
0 Transform + Error JSON, CSV, regex, aggregate, map/filter, error handler, retry, fallback
Explore Workflows API Reference
ORCHESTRATION

Goal pursuit, not
goal tracking.

OKR hierarchy with morning planning and evening review. JARVIS decomposes objectives into daily actions, tracks progress from your screen activity via the awareness pipeline, and escalates when you fall behind.

Get fit by summer 0.45
Run 5K under 25min 0.30
Lose 10 lbs 0.60
Morning run daily
Track calories daily
# Just tell JARVIS your goal
"I want to get fit by summer"

Proposed OKR breakdown:
objective Get fit by summer
key_result Run 5K under 25min
key_result Lose 10 lbs
daily_action Morning run · Track calories
# Awareness feeds goal progress automatically
07:00 Morning plan generated
focus: "Run 5K" is behind pace (0.30/1.0)
action: 30min run scheduled at 6:30am

21:00 Evening review
+0.05 Auto-detected: ran 3.2K via Strava
MISSED: No calorie tracking today
0 Goal Levels objective, key result, milestone, task, daily action
0 = Good Score Google OKR scoring: 0.7 is on track, 1.0 means you aimed too low
0 Daily Rhythm morning plan + evening review, automated and awareness-driven
0 Escalation Stages none → pressure → root cause → suggest kill
Explore Goals API Reference
bash
# 1. Install
$ bun install -g @usejarvis/brain

# 2. Configure
$ jarvis onboard
✓ LLM configured · ✓ Browser found · ✓ Sidecar built

# 3. Launch
$ jarvis start
✓ JARVIS runtime started (PID 4821)
Dashboard: http://localhost:3142

Reachable on

Dashboard :3142 Telegram Discord Gmail Calendar

Your runtime. Your machine. Your agents.

Open source. Self-hosted. Provider-agnostic — bring Anthropic, OpenAI, Gemini, or Ollama.
Full ownership of your data, your memory, and your automations.

Read the Docs Star on GitHub

Contribute

JARVIS is open to contributors.

If you want to improve docs, fix bugs, expand workflows, refine sidecars, or propose new product directions, contributions are welcome. The best starting points are documentation fixes, reproducible bug reports, and focused pull requests against one area of the product.

How to Contribute Open Issues Discuss in Discord