Jarvis โ Autonomous Assistant
Jarvis thinks, plans, and acts on your behalf. in the CLI terminal or the web dashboard. Task queues, proposals, and multi-phase reasoning.
Jarvis is Mode 1 of HelixMind. It is an autonomous AI assistant that runs in the background. planning, analyzing, and executing tasks from a queue. Unlike a regular chatbot, Jarvis does not wait for you to tell it what to do next. It thinks ahead, creates proposals for your review, and executes approved work with multi-phase reasoning.
You can run Jarvis from the CLI terminal or the Web App dashboard. Same engine, same intelligence, two interfaces.
Where does Jarvis run?
| Interface | How to start | Where output appears |
|---|---|---|
| CLI | hx terminal, then /jarvis | Runs as a background session. Output in session tabs. |
| Web App | Jarvis tab in the dashboard | Task panel, proposal queue, real-time output stream. |
Both interfaces share the same task queue. Start Jarvis from the CLI, manage proposals from the Web App. or the other way around.
How Jarvis thinks
Jarvis does not jump to solutions. It reasons in phases, promoting to deeper analysis when the problem demands it.
Quick Check
A fast sweep that takes seconds. Jarvis assesses the current state of your project, checks for obvious issues, and flags anything that needs attention. Think of it as a first glance. enough to catch the low-hanging fruit.
Medium Analysis
When a quick check raises questions, Jarvis goes deeper. It investigates patterns, traces connections between files, and builds a more complete picture. This is where Jarvis starts forming hypotheses about what is happening and why.
Deep Think
For complex problems, Jarvis enters an extended reasoning session with full chain-of-thought analysis. It considers multiple approaches, weighs trade-offs, and constructs a detailed plan of action. Deep Think is the most thorough phase. it takes longer, but produces significantly better results for hard problems.
Each phase can promote to the next automatically. Results from all phases feed into proposals that you review before anything happens.
Tasks
Jarvis works from a task queue. Each task has a title, a description, and a priority level:
| Priority | Behavior |
|---|---|
| Critical | Executed immediately. jumps the queue |
| High | Next in line after any critical tasks |
| Normal | Standard queue order |
| Low | Background work. handled when the queue is clear |
Tasks can come from two sources: you add them explicitly, or Jarvis itself creates them when it discovers issues during analysis. For example, a security scan might generate three new tasks that Jarvis queues up automatically.
Proposals
Before taking any significant action, Jarvis creates a proposal for your review. Each proposal contains:
- What: The specific action Jarvis wants to take
- Why: The rationale behind the decision
- Impact: What will change and where
- Risk: How likely the change is to cause problems
You approve or deny each proposal. When you deny one, your feedback goes back to Jarvis so it can adjust its approach. This keeps you firmly in control. even when Jarvis is running at high autonomy levels, destructive or high-risk actions still surface as proposals.
Autonomy Levels
Control how independently Jarvis operates. You can change the level at any time. raise it when you trust the task, lower it when you want to stay close.
| Level | Name | What Jarvis can do |
|---|---|---|
| L0 | Locked | No tool access. observation only |
| L1 | Read-only | Can read files and analyze, but cannot modify anything |
| L2 | Proposals | Suggests changes, every action requires your approval |
| L3 | Trusted | Safe operations (reads, searches) auto-approved. Writes need approval |
| L4 | Autonomous | Most actions auto-approved. Only destructive operations need approval |
| L5 | Full Autonomy | Everything auto-approved. Jarvis runs completely independently |
Starting at L2 is recommended for most workflows. Move to L3 or L4 once you see Jarvis making good decisions. L5 is for when you fully trust the task and want to walk away.
Workers
Jarvis spawns background workers for parallel execution. When a task can be broken into independent subtasks, Jarvis distributes the work:
- Each worker gets its own session with a separate output stream
- Workers report progress back to the main Jarvis process
- You can monitor all workers in real-time from the CLI tabs or the Web App
- If a worker fails, Jarvis retries the subtask or escalates it for your review
This means Jarvis can refactor three modules simultaneously, run tests in parallel, or analyze different parts of your codebase at the same time.
Using Jarvis in the CLI
Start Jarvis from your interactive terminal:
/jarvis # Start Jarvis
/jarvis add "Refactor auth module" --priority high # Add a task
/jarvis pause # Pause execution
/jarvis resume # Resume
/jarvis status # Show current queue
Jarvis runs as a background session: you keep your interactive prompt and can continue chatting or working. Use Ctrl+PageUp and Ctrl+PageDown to switch between session tabs and see Jarvis output.
Press ESC ESC (double Escape) for an emergency stop that halts Jarvis and all workers immediately.
Using Jarvis in the Web App
Open the Jarvis tab in your dashboard for a visual interface:
- Task queue: See all pending, active, and completed tasks in one list
- Proposal queue: Click Approve or Deny on each proposal with full context
- Autonomy slider: Drag to adjust the autonomy level without typing commands
- Live output: Workers show as real-time output streams below the task panel
- Consciousness events: Watch Jarvis think in real-time as it processes tasks
The Web App connects to your running CLI instance. You can start Jarvis from the CLI and manage everything from the browser, or start it directly from the dashboard.
Brain Integration
Everything Jarvis does is reflected in the 3D Brain visualization:
- Thinking phases appear as animated neuron nodes. yellow and green orbit nodes pulse as Jarvis reasons through problems
- Approved proposals create new task nodes in the helix
- Worker activity generates consciousness events visible in the Brain's live feed
- All knowledge Jarvis acquires is stored in spiral memory, so future sessions benefit from what it learned
Open the Brain (/brain) while Jarvis is running to watch your project's knowledge graph grow in real-time.
Requires Login (Free+)
Jarvis requires a free HelixMind account. Run
helixmind loginin the CLI to unlock it. free forever, no credit card required.