fusefable 0.2.0

Fuse multiple AI models and judge the best answer for coding

Fuse Fable

🌐 Languages: English · ไทย (Thai)

Fan out a coding prompt to many AI models in parallel, then let a judge pick the single best answer — total latency ≈ x2 (slowest model + one judge pass), not x7–x15.

Works three ways: as a CLI, as an MCP server (connect Cursor / VS Code / Claude), and as a subagent / pipe (callable by other tools and scripts).

Install

pip install fusefable            # core
pip install "fusefable[mcp]"     # if you want the MCP server

From source:

git clone https://github.com/proultrax9/fusefable.git
cd fusefable
pip install -e ".[mcp]"

Setup (first run)

fusefable config

• Choose AI Gateway → one key for everything, then it asks "how many models?" and prompts for each one.
  • Known gateways (base URL auto-filled): openrouter, groq, together, fireworks, deepinfra, novita, hyperbolic, aimlapi, portkey, deepseek, openai — any other works too, just type its base URL.
• Or Single providers → it asks how many, then the API kind of each:
  • openai_compat — any OpenAI-compatible endpoint (you provide the base URL)
  • anthropic — Anthropic native (/v1/messages, base URL auto-filled)
  • google — Google Gemini native (generateContent, base URL auto-filled)

Set your API key as an environment variable named as the wizard asks:

export OPENROUTER_API_KEY=sk-...      # macOS/Linux
setx OPENROUTER_API_KEY "sk-..."      # Windows (open a new terminal afterwards)

Config is stored at ~/.fusefable/config.yaml.

1) Use as a CLI

fusefable ask "Write a quicksort function in Python"
fusefable ask --show-all "..."                   # show every answer + judge reason
fusefable ask --models gpt-5,qwen3-coder "..."   # restrict to specific models
fusefable ask --cheap "..."                      # use cheap_models from config
ff ask "..."                                      # short alias

2) Use as a subagent / in a pipe

fusefable ask --quiet "..."                # print only the answer (no headers)
echo "Explain this code" | fusefable ask --quiet   # read the prompt from stdin
cat bug.py | fusefable ask -q "Find the bug in this code"
fusefable ask --json "..."                 # JSON output: answer, chosen_model, reason, cost, candidates

--json is ideal for scripts/agents that parse the result; --quiet is ideal for piping.

3) Use as an MCP server (Cursor / VS Code / Claude / other agents)

Run as an MCP server over stdio:

fusefable mcp

Exposes a tool fuse_ask(question, models?, cheap?) for any MCP client.

Cursor

~/.cursor/mcp.json (or Settings → MCP):

{
  "mcpServers": {
    "fusefable": {
      "command": "fusefable",
      "args": ["mcp"],
      "env": { "OPENROUTER_API_KEY": "sk-..." }
    }
  }
}

VS Code (Copilot / MCP-compatible extension)

.vscode/mcp.json in your project:

{
  "servers": {
    "fusefable": {
      "command": "fusefable",
      "args": ["mcp"],
      "env": { "OPENROUTER_API_KEY": "sk-..." }
    }
  }
}

Claude Desktop

claude_desktop_config.json:

{
  "mcpServers": {
    "fusefable": {
      "command": "fusefable",
      "args": ["mcp"],
      "env": { "OPENROUTER_API_KEY": "sk-..." }
    }
  }
}

Requires pip install "fusefable[mcp]" and a completed fusefable config. If fusefable isn't on the app's PATH, use a full path such as python -m fusefable.cli.

Prompt compression (save tokens)

Reduce token usage while keeping answer quality — useful when you pay per-provider directly. Two tiers, opt-in via --compress:

fusefable ask --compress "<long prompt or pasted code>"
# [compressed: 5200→1800 chars, ~65% saved via llm]

• Tier 1 (lossless): trims trailing whitespace, collapses blank lines, strips zero-width chars — keeps indentation and inner spacing intact (safe for code).
• Tier 2 (LLM): for prompts above compress_min_chars (default 2000), a cheap model compresses semantically — once, then the compressed prompt is sent to all models, so you save tokens × number-of-models.
• Quality guards: prompts under the threshold skip the LLM; if the compressed result is empty, longer, or under 30% of the original, it falls back to the lossless text. The judge always sees the original question.

Config (~/.fusefable/config.yaml): compress, compress_min_chars, compress_model (empty = reuse the judge model).

Architecture

              ENTRYPOINTS (one shared core)
   ┌─────────────┬──────────────────┬─────────────────────┐
   │  CLI        │  pipe / subagent │  MCP server         │
   │ fusefable   │ stdin · --quiet  │ fusefable mcp       │
   │   ask "..." │      · --json    │ tool: fuse_ask()    │
   └──────┬──────┴────────┬─────────┴──────────┬──────────┘
          └───────────────┴────────────────────┘
                          │
                          ▼
                ┌───────────────────┐      ~/.fusefable/config.yaml
                │   core.fuse()     │◀──── (gateway | single providers,
                └─────────┬─────────┘        models, judge, timeout)
                          │
                          ▼
                ┌───────────────────┐
                │  routing          │  build (provider, model) routes
                │  + provider       │  via factory → pick adapter by kind
                │    factory        │
                └─────────┬─────────┘
                          ▼
        ┌──────────────  FAN-OUT (asyncio.gather)  ──────────────┐
        │   ทุกตัวยิงพร้อมกัน · per-model timeout · ตัวพัง = ตัดทิ้ง       │
        │                                                         │
        ▼            ▼              ▼                ▼            ▼
   ┌─────────┐ ┌──────────┐  ┌────────────┐   ┌──────────┐  ┌────────┐
   │ openai_ │ │ anthropic│  │  google    │   │ openai_  │  │  ...   │
   │ compat  │ │ native   │  │  native    │   │ compat   │  │        │
   │(gateway)│ │/v1/msgs  │  │generateCont│   │          │  │        │
   └────┬────┘ └────┬─────┘  └─────┬──────┘   └────┬─────┘  └───┬────┘
        └───────────┴──────────────┴───────────────┴───────────┘
                          │  Completion[] (สำเร็จเท่านั้น)
                          ▼
                ┌───────────────────┐
                │  judge            │  ปกปิดชื่อ → Answer A/B/C...
                │  (anonymized)     │  ให้ judge model เลือกตัวดีสุด
                └─────────┬─────────┘  (พัง → fallback ตัวแรก)
                          ▼
                ┌───────────────────┐
                │  FinalAnswer      │  text · chosen_model · reason
                │  (+ cost estimate)│  · cost_usd · candidates
                └───────────────────┘

Request lifecycle

1. Entrypoint — CLI, a pipe/subagent call, or the MCP tool fuse_ask() — all funnel into one core function core.fuse().
2. Routing — config (gateway or single providers) is turned into (provider, model) routes; a provider factory picks the right adapter per kind (openai_compat / anthropic / google).
3. Fan-out — every model is called concurrently via asyncio.gather (total time = slowest model). Each call has its own timeout; any model that times out or errors is dropped and never slows the run.
4. Judge — model names are anonymized (Answer A/B/C...) so the judge picks on quality alone, not by brand; if the judge fails it falls back to the first answer.
5. Result — returns the best answer with the chosen model, the judge's reason, an estimated cost, and all candidates.

Components (fusefable/)

File	Responsibility
cli.py	Typer CLI (ask / config / mcp), output modes
mcp_server.py	MCP server exposing the fuse_ask tool
core.py	shared fuse() entrypoint + model selection
config.py	load/save config.yaml
wizard.py	interactive setup (gateway vs single, API kind)
routing.py	config → (provider, model) routes
providers/factory.py	pick adapter by kind
providers/openai_compat.py · anthropic.py · google.py	provider adapters
fanout.py	parallel fan-out (drops failures)
judge.py	anonymized judging
fusion.py	orchestrator: fan-out → judge → FinalAnswer
cost.py	token-based cost estimate
models.py	dataclasses: Completion, FinalAnswer, ProviderConfig

Development

pip install -e ".[dev,mcp]"
pytest -q

License

MIT