Fusion

Ask four frontier models at once and get one synthesized answer — locally, through the CLIs you already have logged in. No API keys, no proxy, no token extraction.

Fusion fans a question out in parallel to the locally installed claude, codex, kimi and grok CLIs (each the vendor's own binary, logged in with its normal subscription, running in a sandbox), then has one of them merge the answers into a single Markdown response and say where the models disagreed. Every model answers offline — no web search or browsing — so answers reflect training knowledge, not the live web (why: docs/DESIGN.md). It runs as a small local web app with conversation history; opening it to your LAN or Tailscale network is a one-line opt-in in .env.

                 ┌─ claude ─────┐
question ──┬────►├─ codex ──────┼──► synthesizer ──► answer (streamed) + analysis
           │     ├─ kimi  ──────┤
           │     └─ grok  ──────┘
           └─ earlier fused answers replayed as context

How this compares

The idea — a panel answers, a judge fuses — is the one behind OpenRouter's Fusion router and Karpathy's LLM Council. Fusion differs in where it runs and what it authenticates with.

	Fusion (this repo)	OpenRouter Fusion	llm-council and most API-based councils
Runs	on your machine	in OpenRouter's cloud	on your machine, calls go to OpenRouter
Auth	each vendor's official CLI, already logged in	OpenRouter API key	OpenRouter API key
Cost	your existing subscriptions	sum of every panel and judge completion	metered per token
Models	whatever your claude / codex / kimi / grok are logged into	any OpenRouter slugs	any OpenRouter slugs
Synthesis	one schema'd call: fused answer plus a structured disagreement analysis; candidates anonymized	panel → analyst → outer model, tool-gated (may skip deliberation)	answers → anonymous peer ranking → chairman
Web access	off by design	on for the panel	whatever the models do
Isolation	every CLI in a bubblewrap jail with a scrubbed environment (npm run canary proves it)	n/a	none
History	SQLite; fused answers replayed as context	stateless API	JSON files

Not affiliated with OpenRouter or any of the vendors.

Requirements

• Linux with a Node version that supports node:sqlite and type stripping (see engines in package.json), and bubblewrap (bwrap) — every CLI runs inside a jail that hides the rest of your files from it (see docs/RUNBOOK.md, Security posture).
• Any subset of the four CLIs installed and logged in — untick the ones you don't have. npm run doctor checks.

Quick start

npm install                       # also bundles the frontend libs
cp .env.example .env              # then set the password and cookie secret
npm run doctor                    # CLIs found and logged in?
npm run smoke                     # one trivial call per model
npm start

Open the printed URL, sign in, ask something.

What you see

• Fused answer at the top, streamed as it is written.
• Where the models disagree — consensus, contradictions (and which side the synthesizer took), insights only one model had, gaps all of them missed.
• Raw answers from each model, collapsed, with latency. The synthesizer sees candidates anonymously; names are revealed only here.
• Follow-up questions see the earlier fused answers as context, so all models stay on the same page.

Configuration

Every setting, with its default and meaning, is documented in .env.example.

Running as a service

deploy/fusion.service is a systemd user unit; the install commands are in its header comment. Operations and troubleshooting: docs/RUNBOOK.md.

Terms of service

Fusion stays on the side of the line the vendors draw: it spawns the unmodified official CLI, logged in by you, on your machine, for you. It never reads, stores or forwards an OAuth token, and it does not turn a subscription into an API for anyone else. Anthropic, OpenAI and xAI each document headless use of their own CLI by the subscriber; what they prohibit — and enforce — is lifting the token into another client or re-serving one login to many users. Hence the mandatory password and the single-user design: do not expose this UI to other people on your subscription.

One exception you should know about: Kimi Code's community guidelines reserve its subscription for interactive use. Running that lane from Fusion is your call; the same kimi binary also runs on a pay-as-you-go platform key. Details and sources: the 2026-08-25 entry in docs/DESIGN.md.

Threat model, what the jail does and does not contain, and how to report a vulnerability: SECURITY.md.

Development

npm test (parser tests replay captured CLI output from fixtures/, orchestration tests run against stub providers, plus a docs consistency check), npm run typecheck, npm run fuse -- "question" for a full turn in the terminal, npm run dev for a watching server.

Conventions, the checks a change must pass, and the fixture rule: CONTRIBUTING.md. Design rationale: docs/DESIGN.md. Open work: docs/THREADS.md. Agent guide (for Claude Code and friends): CLAUDE.md.

License

MIT