Quay - Autonomous AI Software Factory

Self-hosted AI agents with parallel ensemble execution, A3M routing, MCP integration, persistent memory, and full cost transparency.

🚀 The Core Differentiator

Quay is the ONLY agent framework with parallel multi-model ensemble execution.

EVERYONE ELSE (litellm, one-api):  try A → fail → try B → fail → try C
Quay (this repo):                   try A + B + C simultaneously → merge results

📦 What's New in v0.5.0

From Vault Research - All Implemented

| Feature | Description | Status | |---------|-------------|--------| | Query-Type Presets | Auto-route by task (code/reasoning/creative) | ✅ | | Cost-Per-Query Tracking | Real-time budget monitoring | ✅ | | Persistent Memory | ReasoningBank 3-phase pipeline | ✅ | | Agent Files | Portable JSON agent configs | ✅ | | MCP Providers | Browser, Bright Data, HuggingFace, Chrome DevTools | ✅ |

🤖 Features

The 5 AI Features

| Feature | Description | |---------|-------------| | Human-in-Loop Verification | Require approval before critical actions | | Determinism & Replay | Black-box recorder for reproducibility | | Visual Pipeline Builder | Drag-drop YAML pipeline generation | | Agent RAG | Knowledge grounding to reduce hallucinations | | Self-Healing Agents | Auto-detect and recover from failures |

The Key Innovation: Parallel Ensemble

// Enable parallel ensemble (THE KEY DIFFERENTIATOR)
const result = await agentRunner.run(task, agent, runId, instructions, tools, {
  useEnsemble: true,
  ensembleConfig: {
    models: [
      { provider: 'openai', model: 'gpt-4o-mini', weight: 0.4 },
      { provider: 'anthropic', model: 'claude-3-haiku', weight: 0.3 },
      { provider: 'google', model: 'gemini-2.5-flash', weight: 0.3 },
    ],
    votingStrategy: 'confidence',
  },
});

Query-Type Presets (P1)

Auto-classifies queries and routes to optimal models:

import { classifyQuery, selectModelForQuery, QUERY_PRESETS } from '$lib/routing';

// Automatic classification
const type = classifyQuery("Write a REST API with authentication");
// → 'code'

// Auto-select optimal model
const { preset, model } = selectModelForQuery(query, { preferSpeed: true });
// → { preset: QUERY_PRESETS.code, model: ModelProfile {...} }

Persistent Memory (P3)

ReasoningBank-inspired memory with 3-phase pipeline:

import { persistentMemory } from '$lib/memory';

// Select relevant memories
const memories = persistentMemory.selectMemory(agentId, query, { limit: 5 });

// Store trajectory and induce memories
persistentMemory.storeTrajectory(trajectory);

// Build context with retrieved memories
const context = persistentMemory.buildContext(agentId, query);

Cost Dashboard

Real-time cost tracking at /cost:

import { costDashboard } from '$lib/features/agentRouter';

// Record a query
costDashboard.record({
  model: 'claude-3-5-sonnet',
  provider: 'anthropic',
  queryType: 'code',
  tokensIn: 500,
  tokensOut: 200,
  costUSD: 0.012,
  latencyMs: 890,
  success: true,
});

// Get metrics
const metrics = costDashboard.getMetrics();

Agent Files

Portable JSON format for agent configuration:

import { AgentFileSerializer, AGENT_TEMPLATES } from '$lib/agentfile';

// Export agent to JSON
const file = AgentFileSerializer.serialize(agentConfig);
await AgentFileSerializer.save(agentConfig, './coder-agent.af');

// Load from template
const coderConfig = AgentFileSerializer.fromJSON(
  JSON.stringify(AGENT_TEMPLATES.coder)
);

// Import
const imported = AgentFileSerializer.load('./coder-agent.af');

🔧 Configuration

# API Keys
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...
QUAY_DEFAULT_LLM_URL=https://api.openai.com/v1

# Budget
QUAY_DAILY_BUDGET=10.00    # Daily cost limit
QUAY_QUERY_BUDGET=0.50      # Per-query cost limit

# MCP Providers
BRIGHTDATA_ZONE=...         # Bright Data for web scraping
HF_TOKEN=...                # HuggingFace access

📁 Project Structure

quay/
├── src/
│   ├── lib/
│   │   ├── ensemble.ts        # Parallel ensemble executor
│   │   ├── dialog.ts          # Multi-round dialog optimizer
│   │   ├── routing.ts         # Query presets + CostTracker
│   │   ├── memory.ts         # PersistentMemory + ReasoningBank
│   │   ├── agentfile.ts      # Agent file format + templates
│   │   ├── mcp-providers.ts  # MCP provider integrations
│   │   └── features/         # 5 AI features
│   ├── server/
│   │   └── agents/
│   │       └── agentRunner.ts
│   └── routes/
│       ├── api/              # API endpoints
│       ├── features/         # Features dashboard
│       ├── cost/             # Cost dashboard
│       └── ...
└── tests/

🌐 Routes

| Route | Description | |-------|-------------| | / | Home / Mission Control | | /features | Features dashboard | | /cost | NEW Cost Dashboard | | /verification | Verification dashboard | | /health | Agent health monitor | | /knowledge | Knowledge base manager | | /pipelines | Visual pipeline builder | | /replay | Determinism replay UI |

🔗 API Endpoints

| Endpoint | Description | |----------|-------------| | GET /api/cost-dashboard | NEW Cost metrics | | POST /api/cost-dashboard | Record query cost | | GET /api/verification | List verification requests | | GET /api/agent-health | Get all agent health status | | GET /api/pipeline-templates | List pipeline templates | | GET /api/knowledge-bases | List knowledge bases | | GET /api/sse | SSE stream |

📊 Test Coverage

Total: 56 tests passing
├── Unit tests: 45
└── Integration tests: 11

📜 License

MIT