Council Mode

⌘J dispatches one prompt to as many LLMs as you wire up, in parallel, then shows you the returns side-by-side with a diff. Vote with 1–6, merge with M, re-dispatch with R.

Why dispatch to multiple LLMs?

Disagreement signals risk. If 5 of 6 models agree but 1 disagrees with a sharp reason, that's a bug worth investigating.
Cost-aware routing. The cheap preset handles routine work; the frontier preset earns its cost on hard problems.
Cross-family tie-break. Claude/GPT/Gemini have different blind spots — diversity beats any single model on average.
Free-tier amortization. The Cheapo preset is $0 forever — you literally cannot pay too much.

Presets

Pre-installed; you can edit them or create your own.

Preset	Voices	Cost	Latency
Quick 3 fast, cheap models. Good for routine code review, rephrasing, doc-gen.	Cerebras gpt-oss-120b · Groq Llama-3.3-70b · Mistral-Large	$0 (all free tiers)	~2–3 s for 3 returns
Frontier Top-shelf reasoning models. For architecture decisions, novel code, gnarly bugs.	Claude Opus 4.7 · GPT-5 · Gemini 2.5 Pro · Cerebras Qwen-3-235B · DeepSeek-V3.1 · Llama-4 Maverick	depends on your paid keys	~5–12 s for 6 returns
Cheapo Maximize free-tier coverage. Burst-friendly.	Cerebras + Groq + Mistral + SambaNova + Gemini AI Studio + OpenRouter free	$0	~3–6 s
Local Zero bytes leave your machine. For sensitive code (legal, regtech, fin).	Ollama qwen2.5-coder · deepseek-r1 · llama-3.3:70b · gemma3 · phi-4 · mistral-small	$0	depends on your GPU
Mix Hand-picked combination. Save your own preset.	any of the above	varies	varies

Graceful degradation

If a voice rate-limits, errors, or times out (40s hard cap), the rest continue. You see a status per voice — green (returned), yellow (in-flight), red (failed). Failed voices don't block the modal.

Voting

Each voice gets a number 1–6 in the order it returns. Press the number to pick that response; press M to merge multiple winners; press Dto see a fine diff between two you can't decide on.

Cost estimator

Before dispatch, the Council modal shows an estimated cost based on prompt length × provider pricing. Local models always show $0. Free-tier models show $0 with the daily quota meter.

Custom dispatch policy

Team+ subscriptions get a config option ~/.furx/dispatch.yml to define your own presets, with rules like:

presets:
  arch-review:
    voices:
      - { provider: anthropic, model: claude-opus-4-7 }
      - { provider: openai, model: gpt-5 }
      - { provider: gemini, model: gemini-2.5-pro }
    require_min_responses: 2
    timeout_ms: 30000
    fallback_to_local: true

Audit

Every Council dispatch writes one row per voice to ~/.furx/furx.db with timestamps, provider, model, latency, token-counts, cost-estimate, and a hash of the prompt. The prompt and response bodies stay local (unless you opt-in to cloud sync).

Provider list — which models work where.
BYOK guide — keys never leave your machine.
Audit log — replay your sessions.