Jarvis Hub — your AI cabinet

Current sprint → Hermes equivalence: status of all 697 capabilities

17 specialized AI agents orchestrated through Jarvis, running on your own hardware, controlled by voice and web.

A private AI OS that cannot silently act beyond the authority you grant it. Local-first and governed: every autonomous action is risk-gated — a reversible/irreversible approval queue for anything consequential, a tamper-evident audit log, budgets and a kill-switch, converging on one Action Kernel mediation point (opt-in via JARVIS_ACTION_KERNEL while it hardens) — and it runs entirely on your own hardware (LM Studio / Ollama on your GPU), $0/month, no cloud by default, with a family agent that never touches the internet. Where other local agent systems prioritize extensibility and raw autonomy, this system prioritizes governed execution, local ownership, and inspectability. The product it is becoming is Nerva (published by Digitaholic; jarvis-hub remains the repo codename until the deliberate rename) — the long-term product & capability vision (house brain, media director, cameras, computer operator, self-extension) is in NERVA_VISION.md.

Get running in minutes → Quickstart · one-click INSTALL.bat on Windows.

---

What this is

A personal AI mesh that handles the cross-section of your life: the day job and the side business, the family, the car and the house projects, fitness, finance, content. Each domain has a dedicated agent. They talk to each other through Jarvis. They learn from interactions.

The agents ship as templates and personalize themselves in your first session: guided onboarding (drop-folder import + profile memory) teaches Jarvis your people, projects, and preferences, every learned fact is inspectable and deletable, and each agent's personality is editable. Your personalized souls live in gitignored agents/<id>/SOUL.local.md overlays — the public repo stays generic, your instance stays yours. The roster below shows each agent's role; the specifics become yours.

Architecture

Pure local Python + opt-in plugin layer. Not OpenClaw — deliberately: where the viral 2026 rival stores secrets in plaintext and runs ungoverned community skills (widely reported in 2026 as a major infostealer target — analysis: docs/research/2026-06-02-personal-ai-competitors.md), here every action is gated by a reversible/irreversible approval queue, guardrails, and an audit log. No cloud dependency by default. Every third-party service is an explicit, scope-limited, audit-able, disable-able plugin.

The cabinet

CNS (Command tier) — 4 agents:

• Jarvis — Prime Orchestrator. The voice you wake.
• Friday — Daily Intel. Weather + news + market signal.
• Pepper — Chief of Staff. Calendar + meetings + reflection.
• Jerome — Leisure & Soundtrack. Music, retro tech, decompression.

Business tier — 5 agents:

• Athena — External Strategist. Your side business, personal brand, career trajectory.
• Stark — Internal Corporate Intel. Day-job KPIs, board prep, channels.
• Veronica — The Voice. Drafts emails, posts, captions in 5 voice profiles.
• Vision — Deep Researcher + OSINT. Cited synthesis, regulatory watch.
• Argus — Geoint Sentinel. Read-only, governed bridge to the Signal Layer.

Tech tier — 3 agents:

• Steve — CTO + Infrastructure. Bonobo + Pi + models + backups.
• Oracle — Pipeline Weaver. n8n workflows, silent when working.
• Ultron — The Shield. Firewall, GDPR/ATT, smart home VLAN.

Foundation tier — 5 agents:

• Gecko — Capital Allocator. Numbers cold; no advice.
• Hercules — Physical Engine. Sleep, recovery, training prep.
• Hephaestus — Builder & Mechanic. Your project car and house build.
• Frigga — The Matriarch. Family memory and care — names, dates, routines. Strictly local-only.
• Howard — Digital Twin (emerging). Voice clone, personal RAG, fine-tune candidate. Local-only.

Hardware

What model should I run? Jarvis works with any LM Studio / Ollama model — pick by VRAM:

Your GPU	Recommended local model	Expectation
8–12 GB (3060/3080)	qwen2.5:7b / llama3.1:8b (Q4)	solid daily chat; deep-think slot off
16 GB (4080/4070Ti S)	qwen2.5:14b / gemma2:27b (Q4)	good quality, one model slot
24 GB+ (3090/4090/5090)	gemma-4-31b-a4b (MoE) + a deep slot	the full reference experience
CPU-only	qwen2.5:3b	it works; expect slow replies

No GPU rule is enforced — the model picker (/api/onboarding/command-center, HUD → Start) tells you honestly whether a model is reachable. The measurement protocol and the per-tier table live in docs/HARDWARE_BENCHMARKS.md — the numbers themselves are not measured yet (owner hardware), and that page stays blank rather than guessing. The reference rig below is what the project is developed on, not a requirement:

• Bonobo WS (Pop!_OS) — Intel Core Ultra 9, 192GB DDR5, RTX 5090 24GB, 4× NVMe + 18TB HDD
• Pi 5 — always-on services (Qdrant, Neo4j, n8n, Homebridge)

Providers & platforms. Any OpenAI-compatible endpoint works, not just LM Studio/Ollama: OpenRouter and other OpenAI-compatible backends ship today (agents/core/llm/openrouter.py, switch with /model). Apple Silicon (M1–M4) is covered — install.sh handles macOS; run local models via LM Studio/Ollama on unified memory (owner smoke-tests tracked as FB1). Running on a lean box or a server/TUI? Use the headless usage profile (JARVIS_SYSTEM_PROFILE=headless) — heavy media features off, background autonomy on, local-light models. Full platform/provider support: docs/COMPATIBILITY.md; tester-level answers: alpha FAQ.

Stack

• Orchestration: Python 3.12 + asyncio + FastAPI
• LLM inference: LM Studio (OpenAI-compatible, primary) → Ollama (fallback) → optional cloud (Anthropic/OpenAI) for approved agents
• Model: google/gemma-4-31b-a4b (MoE, ~4B active params) on RTX 5090
• Voice: browser-mic HUD loop ships; server engines are optional manual installs (openWakeWord wake word, faster-whisper STT, edge-tts / Kokoro TTS) — no install path ships them (docs/VOICE.md)
• Memory: conversation history (JSONL) + vector store (real embeddings via LM Studio/Ollama, hash fallback) + knowledge graph + fused recall (RRF, vector ⊕ graph) + SQLite checkpoints (WAL)
• Channels: web (SSE), voice, telegram, discord, email, slack
• Security: PII/secret scanner, SSRF protection, audit log (Merkle chain), guardrails (WARN/REDACT/BLOCK)

Signal Layer

The Jarvis Signal Layer is the provider-neutral situational-awareness API at :8787. It turns provider data into evidence, signals, relevance, assessments, briefs, and approval-gated recommendations. It starts in deterministic replay mode by default and can later use a WorldMonitor sidecar as provider #1. The WorldMonitor sidecar convention is :3100. See services/signal-layer/README.md.

Run

Generated test matrix: backend 29,663 · frontend 2,202 · mobile 359. Route surface: 707.

One step per OS. Every path ends the same way: a .venv with the hash-pinned dependency set, a passed install smoke (real boot + /readyz + one deterministic chat turn), and the Command Center open at http://127.0.0.1:8080/v2 — bound to loopback only. No cloud key is asked for at any point. Requires Python 3.12+; a local model runtime (Ollama / LM Studio) is detected, never installed.

# Linux / macOS — hosted one-liner (clones into ~/nerva, or $NERVA_DIR)
curl -fsSL https://raw.githubusercontent.com/andrei649/jarvis-hub/main/install.sh | bash

# …or, from a checkout
./install.sh                 # venv + locked deps + smoke, then starts Nerva and opens /v2

Windows 11: double-click INSTALL.bat — it checks for Python 3.12 (offers winget when missing), builds the environment from the lock, and then offers to start Nerva; START.bat launches it later. (PowerShell / packaged executable: packaging\windows\install.ps1 -Launch.)

# Docker, if you would rather not touch the host at all (loopback posture, same as native)
docker compose -f docker-compose.quickstart.yml up --build

# Something not working? One command, named reasons, exit 0/1
python scripts/doctor.py

Full per-OS detail: docs/INSTALL.md. Reaching the HUD from your phone or a second device on the LAN — what the boot guards refuse and the one supported way through: docs/PHONE_ACCESS.md.

Manual, if you prefer to drive it yourself:

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements-beta.txt                 # one install — full feature set
python serve.py              # → http://127.0.0.1:8080
python -m pytest             # full offline suite — current count auto-synced in STATUS.md

Also on Windows: UPDATE.bat pulls the latest and reinstalls. The optional Signal Layer is opt-in on every platform: JARVIS_SIGNAL_LAYER=1 starts it at :8787. If something you added breaks the hub (a persona overlay, a skill, a scheduled job, an MCP server), python serve.py --safe-mode starts it with only what ships: nothing of yours is deleted, the HUD says SAFE MODE, and a normal restart brings it all back.

Changed your mind? ./uninstall.sh --confirm (UNINSTALL.bat on Windows) removes the installer-created Python environment (.venv/). Files outside the current installer footprint are preserved. Your data is never touched unless you also pass --purge-data.

• HUD: http://127.0.0.1:8080/ — the V2 cockpit (primary HUD; legacy HUD at /v1, override with JARVIS_HUD=v1)
• Signal Layer: http://127.0.0.1:8787/healthz (replay mode by default; opt out with JARVIS_SIGNAL_LAYER=0)
• Admin panel: http://127.0.0.1:8080/admin
• CLI REPL: python agents/run.py
• Scripted chat: CLI receipts and session attribution.

Docs

• docs/FEATURES.md — the one-page product sheet: every feature that ships today, how to install, and what's next on the road to 1.0.
• MOONSHOT.md — the north star: vision, principles, phase gates, and how the project stays on track. Read this first for the why and where we're going.
• NERVA_VISION.md — the Nerva product & capability vision: the brand architecture (Cortex/Atlas/Synapse/Vision/Ultron), the six pillars (perception, communication, action, house, media, capability evolution), the target architecture, the capability registry, graduated autonomy, and the measurable "superior to Hermes" bar.
• docs/ARCHITECTURE.md — AI-navigable map of the codebase (entry points, request lifecycle, module index, how-to recipes). Start here to find where things live.
• docs/AI_CONTEXT.md — context-loading map for large-context AI (what to load, in which order, per-task bundles with token estimates).
• docs/OWNER_TASKS.md — the human-gated queue: everything only the owner can do (hardware runs, GitHub settings, decisions).
• docs/RELEASE.md — how a release is cut (tag → source bundle + SBOM + checksums + optional signatures) and how to verify a download.
• SECURITY.md · docs/THREAT_MODEL.md · docs/PRIVACY.md — the trust trio: disclosure policy, what Jarvis defends against (with the mechanism for each threat), and the local-first data/telemetry stance.
• docs/USER_GUIDE.md · docs/FAQ.md · docs/UPGRADE.md — the user docs: install/run/daily-use, quick answers, and how to upgrade safely.
• JARVIS.md — architecture & directory structure · AGENTS.md — assistant conventions · BACKLOG.md — priorities & tasks.
• GO_LIVE_PLAN.md — features + marketing brief + v1.0 launch checklist · docs/VALUATION_AND_PRICING.md — valuation, pricing & unit economics.
• docs/MANUAL_TESTING.md — human pre-release checklist: everything the offline test suite can't verify (real LLMs, channels, services, HUD rendering).
• — forward roadmap: remaining v1.0 gate, WorldView follow-ups, audit-debt hardening, post-1.0 horizons (Hermes, Cognition), and recommended sequencing.

Status

Generated status: v1.0.0 · backend 29,663 · frontend 2,202 · mobile 359 · 707 routes · 18 active agents · open release gates: A1, A3, A4, A6 · backlog: 289 done · 21 open or blocked of 310 horizon rows (proof status: docs/OWNER_TASKS.md → Production-verification checklist) · source commit 01ceee236e90. Full data: project-status.json.

v1.0.0 — the proof track is closed (the version line is the roadmap — see BACKLOG.md: productionization done, the 72h soak now grades itself, design partners running on non-owner installs, owner-host proof passed. The AI-OS capability program, NERVA_VISION.md, continues past 1.0 as roadmap rather than as a gate). 18 specialist agents (incl. Argus for evidence-based situational awareness, Howard, the emerging digital twin, and Hestia, the House Brain; + 14 bench) across 4 tiers; real-embeddings recall (LM Studio) + fused recall + RAG injection; hot-path perf (SQLite WAL, event-loop offload, checkpoint debounce, query-embedding cache, complexity-based model tiering); autonomous proactive cortex (ORIZONT 6); security wedge (encrypted secrets, signed skills, reversible/irreversible approval split, quarantine/capability/kill-switch); competitive edge (workflow engine, model arena, quality monitor, review queue); living memory (bi-temporal KG, decay-forgetting, sleep-time consolidation). Full offline test suite green — current counts auto-synced in STATUS.md (+ frontend JS tests).

Road to v1.0: the original feature backlog (H1–H17) is code-complete at 194/196 items (≈99% by story points) — the only two open items (H12.14 fine-tuned agentic model, H13.3 speculative decoding) need the GPU host (runbook: docs/GPU_RUNBOOK.md). What stands between 0.11.0 and the 1.0.0 tag is now two-fold: (a) the proof track — the productionization layer (H23: agentic-safety budgets, DB migrations, backup/restore + export-delete, operability, quality + user docs — see the version roadmap) proven with real design-partner users — and (b) the AI-OS capability program (ORIZONT 27–33: capability registry, computer/browser operator, media director, house brain, cameras, capability acquisition, ambient intelligence — NERVA_VISION.md, gate expanded 2026-07-11); manual testing (docs/MANUAL_TESTING.md) + code audit (docs/AUDIT.md) are the release step that tags a version. The HUD V2 cockpit is the default UI; deep write-controls for ~37 newer backend surfaces are tracked in docs/design/HUD_V2_REMAINING.md. See BACKLOG.md + MOONSHOT.md §4.

See STATUS.md, BACKLOG.md, and docs/ARCHITECTURE.md for details.