Capafy logo Capafy
开始使用
Multi-Model AI Agent Starter Kit

Multi-Model AI Agent Starter Kit

by: Erkan
One-call orchestration across OpenAI, Anthropic, and Google Gemini with role assignment, voting, cost-aware routing, and a unified response envelope. Ships with LangChain and CrewAI adapters, an offline mock fallback for testing without API keys, and a 100-point quality/cost/latency score.
#工程#研究#分析
评分
需要更多评价
订单数
0
如何使用
下载

🧠 Multi-Model AI Agent Starter Kit

Orchestrate GPT + Claude + Gemini in one call.

Python 3.10+
99 tests
stdlib only
License: MIT

Run, compare, and ensemble multiple LLM providers in a single API call.
6 orchestration strategies, streaming, circuit breaker, hallucination
self-check, decision receipt (audit trail), output boundary guard,
batch runner, and a full CLI included.

✨ What sets it apart (11 features competitors lack)

# Feature Why it matters
1 Streaming providers Chunked output from any provider
2 Per-provider circuit breaker Auto-failover on cascading failures
3 Exponential backoff retry + jitter Survives 429/503/timeout without crashing
4 Hallucination self-check Citation density + numeric consistency + length sanity
5 Decision receipt (SHA-256) Full audit trail, JSON-exportable
6 Output boundary guard Auto-disclaimer for legal/medical/financial/compliance tasks
7 CLI run / compare / stream / doctor
8 Batch runner N tasks, shared cost cap, parallel
9 LRU + TTL response cache Hit-rate stats, shareable across runs
10 Red-team strategy Generator + critic with explicit flaw injection
11 Pure Python stdlib core Zero required runtime dependencies

🚀 Quick start

pip install multi-model-agent-starter

import asyncio
from MULTI_MODEL_AGENT_STARTER import run_agent_team

async def main():
    result = await run_agent_team(
        task="Write a positioning statement for an AI note-taking SaaS",
        agents=[
            {"role": "strategist",  "provider": "openai",    "model": "gpt-4o-mini"},
            {"role": "critic",      "provider": "anthropic", "model": "claude-3-5-haiku"},
            {"role": "factchecker", "provider": "google",    "model": "gemini-1.5-flash"},
        ],
        strategy="vote",
    )
    print(result["final_answer"])
    print(result["quality"])
    print(result["receipt"]["content_hash"])  # SHA-256 audit hash

asyncio.run(main())

🧰 CLI

# Offline default (mock provider)
python -m MULTI_MODEL_AGENT_STARTER run "Write a tagline for a calculator" --strategy vote
python -m MULTI_MODEL_AGENT_STARTER compare "Explain RAG" --providers mock,mock
python -m MULTI_MODEL_AGENT_STARTER stream "Tagline" --provider mock
python -m MULTI_MODEL_AGENT_STARTER doctor

📊 Strategies

Strategy Calls When to use
single 1 Cheapest path with one provider
vote N (parallel) Independent judgments, then majority answer
debate 2N + 1 Nuanced tasks — critiques before synthesis
router 1 (cost-tier matched) Cost-sensitive, complexity-aware
chain N (sequential) Each step builds on the prior output
red_team 2N rounds Adversarial QA — surfaces flaws before commit

🔍 Output envelope

{
  "task": "Write a positioning statement",
  "task_type": "writing",
  "complexity": "medium",
  "strategy": "vote",
  "final_answer": "...",
  "quality":     {"total": 86, "coverage": 22, "clarity": 23, "completeness": 25, "efficiency": 16},
  "hallucination": {"risk_score": 0.12, "citation_density": 0.6, "numeric_consistency": 1.0},
  "boundary":     {"is_high_risk": false, "flagged_phrases": []},
  "cache":        {"hits": 1, "misses": 2, "hit_rate": 0.33},
  "breakers":     {"openai": {"state": "closed"}},
  "receipt":      {"run_id": "638d30bc-...", "content_hash": "9d4d4950eb03c95f..."}
}

📦 Modules (16)

core/task_analyzer · core/role_assigner · core/boundary_guard ·
providers/{base,registry,mock_provider,openai_provider,anthropic_provider,google_provider} ·
providers/{circuit_breaker,retry,cache,streaming,hallucination_checker,pricing} ·
orchestrator/{strategies,cost_tracker,orchestrator,batch_runner,receipt} ·
scoring/quality_scorer · formatter/output_formatter · __main__ (CLI)

✅ Tests

99 passed in 0.72s

pytest tests/ — unit + integration markers. Runs offline, no API key
required.

🔌 Optional dependencies

pip install openai>=1.30          # OpenAI provider
pip install anthropic>=0.30       # Anthropic provider
pip install google-generativeai>=0.5  # Gemini provider
pip install langchain-core>=0.2   # LangChain adapter
pip install crewai>=0.30          # CrewAI adapter

Nothing required for the core — mock_provider ships in the box.

⚠️ Honest limitations

  • Token counts are estimated when providers don't return usage
  • Pricing is list price — override with negotiated rates if needed
  • Mock provider is deterministic — useful for tests, not for production
  • No tool calling wired in yet — text orchestration only

💰 Pricing

Download (one-time): $11 — indie devs, AI builders, agencies.
Full source, all 16 modules, 99 tests, 4 examples, CLI included.
6 months of free updates.


View more