多模型AI编排服务器

graph TB
    subgraph Claude Code
        CC[Claude Code / MCP Client]
    end

    subgraph "AarnXen MCP Server (ax)"
        CC -->|MCP Protocol| MW[Tool Middleware]
        MW -->|rate limit, guardrails, events| TOOLS

        subgraph TOOLS["31 Tools"]
            direction LR
            AI["AI Tools<br/>chat, consensus, compare<br/>think, codereview, precommit<br/>challenge, swarm, pipeline<br/>debate, verify, jury, refine"]
            WEB["Web Tools<br/>web_search<br/>web_fetch"]
            KB["KB Tools<br/>kb_store, kb_search<br/>kb_get, kb_remember<br/>kb_recall, kb_relate<br/>kb_search_index<br/>kb_timeline<br/>kb_get_observations<br/>kb_stats"]
            SYS["System Tools<br/>costs, list_models<br/>provider_health<br/>health, events<br/>kb_consolidate"]
        end

        TOOLS --> ROUTER[Smart Router]
        ROUTER -->|tier selection| REG[Provider Registry]
        ROUTER <-.->|learns from| KBD[(Knowledge DB)]

        subgraph Providers
            direction LR
            GEM[Gemini API<br/>2.5 Pro, 2.5 Flash<br/>3.1 Pro, 3 Flash]
            OLL[Ollama Cloud<br/>DeepSeek, Kimi<br/>Qwen, GLM, MiniMax<br/>GPT-OSS, Mistral<br/>Cogito, Devstral]
        end

        REG --> GEM
        REG --> OLL

        subgraph Infrastructure
            direction LR
            CACHE[Response Cache<br/>TTL + LRU]
            CB[Circuit Breaker<br/>3-state]
            COST[Cost Tracker<br/>per-request USD]
            EVENTS[Event Bus<br/>async pub/sub]
            GUARD[Guardrails<br/>content filter]
            RATE[Rate Limiter<br/>per-tool]
        end
    end

    style TOOLS fill:#1a1a2e,stroke:#16213e,color:#fff
    style Providers fill:#0f3460,stroke:#16213e,color:#fff
    style Infrastructure fill:#533483,stroke:#16213e,color:#fff

graph LR
    A[AI Tool Called] -->|auto-learn middleware| B[Store Performance<br/>latency, tokens, cost]
    B --> C[(Knowledge DB<br/>SQLite + FTS5)]
    C -->|memory-informed routing| D[Smart Router<br/>reorders by real perf]
    D -->|picks fastest model| A

    style A fill:#e94560,stroke:#16213e,color:#fff
    style B fill:#533483,stroke:#16213e,color:#fff
    style C fill:#0f3460,stroke:#16213e,color:#fff
    style D fill:#1a1a2e,stroke:#16213e,color:#fff

# Clone
git clone https://github.com/Abhishek-Kraj/AarnXen-MultiAI-MCP.git
cd AarnXen-MultiAI-MCP

# Install (pick one)
uv sync                              # recommended
pip install -e .                      # alternative

# Configure
mkdir -p ~/.aarnxen
cp config.example.yaml ~/.aarnxen/config.yaml
# Edit config.yaml — add your API keys

# Verify
uv run pytest tests/ -q              # 269 tests should pass

claude mcp add ax -s user -- uv --directory /path/to/AarnXen-MultiAI-MCP run ax

{
  "mcpServers": {
    "ax": {
      "command": "uv",
      "args": ["--directory", "/path/to/AarnXen-MultiAI-MCP", "run", "ax"],
      "env": {
        "GEMINI_API_KEY": "your-gemini-key",
        "OLLAMA_CLOUD_KEY": "your-ollama-cloud-key"
      }
    }
  }
}

graph LR
    subgraph Budget["💰 Budget Tier"]
        B1[gemini-3.1-flash-lite]
        B2[gemini-3-flash]
        B3[gpt-oss:20b]
        B4[qwen3.5:397b]
    end

    subgraph Balanced["⚖️ Balanced Tier"]
        M1[kimi-k2:1t]
        M2[deepseek-v3.2]
        M3[glm-5]
        M4[mistral-large-3:675b]
    end

    subgraph Premium["👑 Premium Tier"]
        P1[gemini-3.1-pro]
        P2[kimi-k2-thinking]
        P3[minimax-m2.5]
        P4[cogito-2.1:671b]
    end

    Q[Query] --> R{Smart Router}
    R -->|simple task| Budget
    R -->|general task| Balanced
    R -->|complex reasoning| Premium

    style Budget fill:#2d6a4f,stroke:#1b4332,color:#fff
    style Balanced fill:#e85d04,stroke:#dc2f02,color:#fff
    style Premium fill:#7b2cbf,stroke:#5a189a,color:#fff

"Use ax chat to ask gemini-2.5-flash about Python async patterns"

"Use ax consensus to get opinions from 3 models on microservices vs monolith"

"Use ax debate with topic='Is Rust better than Go for backend services?' rounds=3"

"Use ax verify with claim='Python is the most popular programming language' web_check=true"

"Use ax jury to score this code with criteria=code and num_jurors=5"

"Use ax refine with prompt='Write a Redis connection pool in Python' iterations=3"

"Use ax codereview with focus=security on this authentication handler"

"Use ax precommit to review my staged changes"

"Use ax web_search for 'best practices for MCP server development'"
"Use ax web_fetch url='https://docs.anthropic.com/en/docs/agents-and-tools/mcp'"

"Use ax kb_remember 'Always use connection pooling for PostgreSQL in production'"
"Use ax kb_recall 'PostgreSQL'"
"Use ax kb_search 'database optimization'"

"Use ax pipeline with steps:
  1. web_search for 'Python GIL changes in 3.13'
  2. think about the implications
  3. codereview on our threading code"

graph TD
    subgraph "Single Model"
        chat[chat]
        think[think]
        codereview[codereview]
    end

    subgraph "Multi-Model Parallel"
        consensus[consensus]
        jury[jury]
        swarm[swarm]
    end

    subgraph "Multi-Model Sequential"
        debate[debate<br/>FOR ↔ AGAINST → Judge]
        verify[verify<br/>Draft → Questions → Answers → Verdict]
        refine[refine<br/>Generate → Critique → Refine × N]
        pipeline[pipeline<br/>Tool₁ → Tool₂ → Tool₃]
    end

    subgraph "Web Grounding"
        web_search[web_search]
        web_fetch[web_fetch]
    end

    verify -.->|optional web_check| web_search

    style chat fill:#2d6a4f,color:#fff
    style think fill:#2d6a4f,color:#fff
    style codereview fill:#2d6a4f,color:#fff
    style consensus fill:#e85d04,color:#fff
    style jury fill:#e85d04,color:#fff
    style swarm fill:#e85d04,color:#fff
    style debate fill:#7b2cbf,color:#fff
    style verify fill:#7b2cbf,color:#fff
    style refine fill:#7b2cbf,color:#fff
    style pipeline fill:#7b2cbf,color:#fff
    style web_search fill:#1d3557,color:#fff
    style web_fetch fill:#1d3557,color:#fff

default_model: "auto"
default_temperature: 0.7

providers:
  - name: gemini
    api_key_env: GEMINI_API_KEY
    priority: 1
    models:
      - gemini-2.5-pro
      - gemini-2.5-flash
      - gemini-3.1-pro-preview
      - gemini-3-flash-preview

cache:
  enabled: true
  ttl_seconds: 3600

memory:
  enabled: true
  path: "~/.aarnxen/conversations.db"

cost_tracking: true

providers:
  - name: gemini
    api_key_env: GEMINI_API_KEY
    priority: 1
    models:
      - gemini-3.1-pro-preview
      - gemini-3-flash-preview
      - gemini-2.5-pro
      - gemini-2.5-flash
      - gemini-2.5-flash-lite

  - name: ollama-cloud
    base_url: "https://ollama.com"
    api_key_env: OLLAMA_CLOUD_KEY
    priority: 2
    models:
      - deepseek-v3.2
      - kimi-k2:1t-cloud
      - kimi-k2.5
      - kimi-k2-thinking
      - qwen3.5:397b-cloud
      - glm-5
      - minimax-m2.5
      # ... see config.example.yaml for full list

http://localhost:8765

sequenceDiagram
    participant C as Claude Code
    participant MW as Middleware
    participant RL as Rate Limiter
    participant GR as Guardrails
    participant T as Tool Handler
    participant AL as Auto-Learn
    participant EB as Event Bus

    C->>MW: tool call
    MW->>RL: check rate limit
    RL-->>MW: ✅ allowed
    MW->>GR: validate input
    GR-->>MW: ✅ safe
    MW->>EB: emit tool.start
    MW->>T: execute handler
    T-->>MW: result
    MW->>AL: store model performance
    MW->>EB: emit tool.complete
    MW-->>C: response

CLOSED → (3 failures) → OPEN → (30s cooldown) → HALF_OPEN → (1 success) → CLOSED
                                                            → (1 failure) → OPEN

# Install dev dependencies
uv sync --group dev

# Run tests
uv run pytest tests/ -q                    # all 269 tests
uv run pytest tests/test_router.py -q      # specific module
uv run pytest tests/ -q -x                 # stop on first failure

# Run server locally
uv run ax

# Run dashboard standalone
uv run aarnxen-dashboard

src/aarnxen/
├── server.py              # FastMCP server + tool registration
├── config.py              # YAML config loader
├── dashboard.py           # Web dashboard (auto-starts)
├── core/
│   ├── router.py          # Smart 3-tier routing + memory-informed
│   ├── tool_middleware.py  # Rate limit, guardrails, events, auto-learn
│   ├── knowledge.py       # SQLite + FTS5 knowledge base
│   ├── circuit_breaker.py # 3-state circuit breaker
│   ├── events.py          # Async event bus
│   ├── guardrails.py      # Content filtering
│   ├── rate_limit.py      # Per-tool rate limiting
│   ├── cost.py            # USD cost tracking
│   ├── cache.py           # TTL + LRU response cache
│   ├── retry.py           # Exponential backoff + fallback
│   ├── conversation.py    # Persistent conversation memory
│   ├── extractor.py       # Entity extraction
│   └── validation.py      # Input validation
├── providers/
│   ├── registry.py        # Provider registry + model resolution
│   ├── base.py            # Base provider interface
│   ├── gemini.py          # Google Gemini (native SDK)
│   ├── ollama.py          # Ollama Local + Cloud
│   └── openai_compat.py   # OpenAI-compatible endpoints
├── tools/
│   ├── chat.py            # Single model chat
│   ├── consensus.py       # Parallel multi-model
│   ├── compare.py         # A/B comparison
│   ├── think.py           # Deep reasoning
│   ├── codereview.py      # Code review
│   ├── precommit.py       # Pre-commit gate
│   ├── challenge.py       # Devil's advocate
│   ├── debate.py          # Adversarial debate
│   ├── verify.py          # Chain of Verification
│   ├── jury.py            # N-model jury voting
│   ├── refine.py          # Self-Refine
│   ├── web_search.py      # DuckDuckGo search
│   ├── web_fetch.py       # URL → markdown
│   ├── swarm.py           # Parallel sub-tasks
│   └── pipeline.py        # Sequential tool chain
└── pricing/
    └── models.py          # Per-model pricing data