INITIALIZING NEURAL CORE
0%
_
J.A.R.V.I.S. v2.0 · LOCAL
DOWNLOAD

Conversation

LOCAL LLM · LLAMA 3.x

Settings

✓ SAVED LOCALLY
TS
Sir Stark
sir@starkindustries.com
LOCAL ACCOUNT · ENCRYPTED

Identity

Jarvis greets you by name and signs drafts as you.

Account & security

Used for connectors, drafts and account recovery. Credentials are stored in the system keychain — never in plaintext.

Change password

Assistant

Shapes how Jarvis talks and behaves across every conversation.

Prepended to every conversation. Tool instructions stay in their own definitions.

Long-term memory

Durable facts Jarvis keeps about you — names, preferences, recurring projects — injected into every conversation so it remembers across sessions. Jarvis adds these itself when you share something worth keeping; you can add or remove them here.

No memories yet — tell Jarvis something worth remembering.

OpenAI

The whole AI surface runs on one OpenAI key. It's held only by the local core (OS keychain) — never stored in the app bundle or sent anywhere but OpenAI, on your key.

Checking… · leave blank to keep the current key; clear and save to remove it.

Usage & cost

Live tally of OpenAI tokens and estimated spend on your key — counted from the actual API responses. Set a monthly budget and Jarvis warns you as you approach it.

$0.00
This month
0
Input tokens
0
Output tokens
0
Requests
No budget set — usage is tracked but not capped.
All-time
No usage yet.

Computer control

Lets Jarvis — and any agent you grant the Computer / shell capability — act on this computer: open apps and folders, find files, run builds, deploy. Enabled by default, but nothing runs on its own: every command shows up as a card (in chat or in the agent's run window) and executes only after you click RUN. This toggle is the kill-switch — flip it off and execution stops instantly, everywhere, including agents.

Allow Jarvis to run commands
One command at a time · your approval required for each
Checking…
Audit log — every executed command
Nothing executed yet.

Setup guide

Everything you need to configure for Jarvis to be fully operational.

  1. OpenAI API key — required. The whole AI surface (chat, voice, tools, agents) runs on one key. Create it at platform.openai.com/api-keys (account needs billing enabled — pay-as-you-go, a few $ is plenty to start), paste it in the OpenAI section above and hit SAVE CHANGES. It goes straight to your OS keychain — never stored in the app or sent anywhere except OpenAI.
  2. Microphone. Press the mic button once — the OS will ask for permission on first use (macOS: System Settings → Privacy → Microphone; Windows prompts automatically). After that voice is hands-free: just talk, a ~1.3s pause ends your turn, Jarvis answers out loud and keeps listening. Press the mic again to stop.
  3. Identity & persona. Set your name and tune the core system prompt in the sections above — it shapes how Jarvis talks in every conversation and for every agent.
  4. Computer control — on by default. Jarvis can operate this machine, but never silently: every command appears as a card in chat and runs only after you click RUN; everything lands in the audit log. The toggle above turns the whole capability off instantly.
  5. Updates. The app checks GitHub Releases for new versions. You can always grab the latest installer from the project's Releases page.

Docs — how Jarvis works

The full picture, plainly. Click a topic to expand.

Architecture & privacy

Jarvis is a native desktop app (macOS & Windows). The interface you see is a web layer; underneath runs a compiled Rust core that does all the real work and owns your OpenAI key.

  • The key lives only in the OS keychain (macOS Keychain / Windows Credential Manager) — never in the app bundle, never in logs, never sent anywhere but OpenAI on your own account.
  • The web layer talks to the core through small typed calls and receives streamed results back — it never touches the key or calls OpenAI directly.
  • Everything is local-first: your tasks, notes, finance, memory and audit log stay on your machine.
Chat, voice & memory
  • Chat streams replies token-by-token from gpt-4o-mini. Transient network/rate errors retry automatically with backoff.
  • Voice is hands-free: press the mic once, just talk. Local voice-activity detection ends your turn after ~1.3s of silence, transcribes it, Jarvis answers and speaks each sentence as it's generated, then keeps listening. Start talking over it to interrupt (barge-in).
  • Short-term memory — the recent back-and-forth of the current conversation is always in context.
  • Long-term memory — durable facts (names, preferences, projects) are saved with the memory_save tool and injected into every future conversation, so Jarvis remembers across sessions and restarts. Manage them in the Long-term memory section above.
  • Conversation archive — every chat is saved in full. The clock icon in the chat header opens the history: browse all past conversations, search across them by any word, reopen one to keep talking, delete, or export them all to JSON. The + starts a fresh conversation.
Skills & live panels

Jarvis acts through tools. When you ask for something, it calls the right one and the matching panel updates live:

  • Tasks/boards — add, move, list, create boards.
  • Goals & tracker — set goal progress, log habits.
  • Knowledge vault — save and search notes.
  • Finance — record expenses, summarize the budget.
  • Research — queue a topic to search & summarize.
  • Memory — remember/forget durable facts.
Computer control & agents

Jarvis — and any agent you grant the Computer / shell capability — can operate this machine: open apps and folders, find files, run builds, deploy.

  • Nothing runs on its own. A command appears as a card showing the exact text; it executes only after you click RUN (30s timeout, output shown inline).
  • Every run and denial is written to the audit log (in Computer control), tagged with the agent name when an agent issued it.
  • The Computer control toggle is the kill-switch — turn it off and execution stops instantly, everywhere, including agents.
  • Agents are scoped assistants you build with their own prompt, tools and knowledge. They run in their own window with their own memory.
Cost control & updates
  • Usage & cost tallies real tokens from every response and estimates spend per month. Set a monthly budget and Jarvis warns you as you approach it.
  • Updates — the app checks for new signed releases and offers a one-click “Update & restart”. Each update's signature is verified before it installs.

Privacy

Local-first by default. Cloud features are opt-in and run on your own keys.

Local-only mode
Audio and data never leave this device
Cloud LLM fallback
Use a cloud model when local is unavailable
Telemetry
Anonymous usage analytics