JharvisAgent 0.1.0 — lightweight local-first GitHub base (runtime source only).
Ask one question. Several models answer. You see exactly where they disagree.
A council of models that runs on your own machine — free, offline if you want it, and it learns which model is actually good at your work.
- What it does
- Install in 3 minutes
- First run
- The five screens
- Your keys stay on your machine
- Full documentation
- Roadmap
- License
| Desk | Side-by-side answers from multiple models on one prompt. |
| Debate | Parallel agent rounds — answer, anonymised peer critique, revise — with an optional consensus gate before synthesis. |
| Divergence | Claim-level view of where the models agree and where they split. |
| Office | Hired character agents with persistent missions, refusals, themes and model lanes. |
| Hall of Fame | Per-model, per-task-type scores built from your own runs and your own blind picks. |
Runs free and offline. Ollama is the default. Offline mode refuses every cloud provider at the provider layer — not just hidden in the UI.
The agents persist. Each hire is staff with a mission, explicit refusals, and things it must ask for before starting.
It learns your work. Scores come from your prompts and your blind taste tests, not a public leaderboard.
Not a voice assistant. Not a task router. Not a personal assistant. Not production-grade yet.
Requirements: Node.js 20 or newer. Ollama if you want it free and offline (recommended).
git clone https://github.com/MLDCode/jharvisagents.git
cd jharvisagents
cp .env.example .env # Windows: copy .env.example .env
npm install
npm run devOpen http://localhost:3456.
Then pull at least one local model so the council has something to run:
ollama pull llama3.2Production check: npm run build && npm start.
Full walkthrough, per-OS notes and troubleshooting: docs/INSTALL.md
Quick Setup opens automatically and walks four steps: You → Office → Models → Assign.
- You — your name, UI mode (Jharvis HUD or plain), light/dark, seasonal theme.
- Office — office type and which agent roles to hire.
- Models — check Ollama, start it if it isn't running, paste optional cloud keys.
- Assign — give each agent a provider and model.
You can skip it and configure everything later in Settings.
Ask your first question from the prompt bar at the bottom of the Desk. Select up to four agents on the roster, then Run debate.
Every screen explained, plus how to customise the UI: docs/USAGE.md
The HUD workspace. System status on the left, agent roster on the right, debate configuration along the bottom: how many agents, critique / jury / diff mode, judge mode, rounds, and the consensus threshold. Cost and time estimate before you commit.
Arm a prompt, pick agents, fire once. Each card tracks queued → working → done / fail
and uses that agent's own provider and model.
Tower or floor view of every hired agent, their live log line, and their status. Click an agent to open its detail and chat.
Manual, interval, daily or weekly schedules resolved in a real IANA timezone. Slots missed while the app was closed run once on reopen.
Seven providers: Ollama, Groq, Gemini, OpenRouter, OpenAI, Anthropic, Kimi. Offline-only lock, paid-model visibility, per-provider keys — stored on this machine only and shown masked once set.
API keys live in data/settings.json and .env. Both are gitignored. There is no
cloud backend for this app — nothing is sent anywhere except directly to the
provider you configured, from your machine.
The settings API never returns a key. It returns only configured: true/false and a
masked hint.
Before you share a clone or open a PR:
npm run reset # clears keys and local agent/office/task artefacts
npm run check:secrets # fails if any tracked file or commit looks like a credentialTurn on Offline only in Settings and every cloud provider is refused at the provider layer — requests never leave the process.
| Document | What's in it |
|---|---|
| docs/INSTALL.md | Download, install, per-OS notes, Ollama setup, cloud keys, upgrading, troubleshooting. |
| docs/USAGE.md | Every screen, the debate pipeline, scheduling, council packs — and how to customise or extend the UI. |
| docs/WIKI.md | Single-file architecture and API reference. |
| docs/wiki/ | Per-topic wiki pages (also readable in-app at /wiki). |
| CHANGELOG.md | What shipped, by version. |
| docs/github-setup.md | Suggested repo description and topics. |
Debate depth. Desk rounds (1 / 2 / 3) run R1 answers, optional R2 anonymised peer critique, and optional R3 revise before consensus gate / judge.
Optional modules. API TTS, a local browser-agent sidecar, ComfyUI image/video, and an in-app GGUF picker stay out of core — use Ollama or LM Studio for local weights.
Ship shape. Harden for longer unattended use, live Hall refresh after debates, and the remaining Apps tools still marked later.