🔀 Fuse

An internal Mixture-of-Agents chat tool. Ask one question → several AI models (Claude, GPT, Grok, Gemini, or anything via OpenRouter) answer in parallel → one "aggregator" model fuses their answers into a single best response. Tracks token usage locally with charts.

Stack

• Next.js 14 (App Router) - frontend and backend (API routes). One process.
• React + Tailwind UI, Recharts for stats.
• Electron wrapper → ships as a double-clickable Fuse.app (no terminal).
• Provider adapters for Anthropic, OpenAI, xAI, Google, and OpenRouter.
• Token usage and full chat history stored locally as JSON (see Data below).

Install with your coding agent (easiest)

Already have Claude Code or Codex CLI? Paste this prompt into it and it will set Fuse up and launch it for you - no manual steps:

Install and launch the Fuse desktop app on my Mac from its public repo. Do this:

1. Ensure git and Node.js 18+ are available. If Node is missing, install it with Homebrew: `brew install node`.
2. Clone https://github.com/LiamBCN/fuse.git into ~/fuse and cd in (git pull if it already exists).
3. Run `npm install`.
4. Build the native app: `npm run pack`. This produces Fuse.app under dist/ (e.g. dist/mac-arm64/Fuse.app).
5. Copy that Fuse.app into /Applications and open it with `open -a Fuse`. A native window should appear.
6. If the native build fails for any reason, fall back to `npm run dev` and open http://localhost:3030 instead.

Fuse routes answers through my local Claude and Codex CLIs, so once it's open, go to Settings to confirm the detected providers and pick proposer + aggregator models. Tell me when it's running.

Fuse detects your installed CLIs on first launch and walks you through the rest.

Run as a Mac app (recommended)

npm install
npm run dist        # builds dist/Fuse-<version>-arm64.dmg

Open the .dmg, drag Fuse to Applications, launch it. It starts its own server internally and opens a native window - nothing to run in a terminal. (The app is ad-hoc signed, not notarized, so the first launch may need right-click → Open to get past Gatekeeper.)

Run in the browser (dev)

npm install
npm run dev            # http://localhost:3030
# or: npm run electron:dev   - dev server + Electron window with live reload

1. Go to Settings → paste the API keys for whichever providers you want.
2. Pick your proposer models (they all answer) and one aggregator (it fuses).
3. Chat. Expand "show proposals" or hit Debug ↗ on any answer to see what each model said.
4. History tab lists every conversation; open one to debug each agent's answer. Resume reopens it as the live chat.
5. Stats tab shows tokens/day, tokens by model, and estimated cost.

How it works

your question
   ├─► Claude  ─┐
   ├─► GPT     ─┤   (proposers, in parallel)
   ├─► Grok    ─┤
   └─► Gemini  ─┘
                 └─► Aggregator model ──► final fused answer

This is the Mixture-of-Agents recipe (Together AI, ICLR 2025): models write better answers when they can see other models' answers first.

Data

Config (API keys + model choices), conversations, and token usage are stored as JSON on disk:

• Packaged app: ~/Library/Application Support/fuse/data/ - settings.json, conversations.json, usage.json, plus port.json (the app reuses one stable local port so its origin/localStorage stay consistent across launches).
• Dev (npm run dev): ./data/ in the repo.
• Override the location with the FUSE_DATA_DIR env var.

Everything is durable across restarts and refreshes. A legacy config in browser localStorage is auto-migrated into settings.json on first load.

Notes

• API keys are stored locally on this machine in settings.json and sent per-request to the local server. They are never sent anywhere except the model providers you configured.
• Pricing in lib/models.ts is best-effort and editable; unknown models still track tokens, just with a $0 cost estimate.
• Add new models by typing any model id in Settings (free text + suggestions).