GitHub - Ajayduddi/AutoPilot: AutoPilot transforms how you interact with automation workflows. Instead of manual triggers and complex dashboards, simply chat with your automations. It connects to multiple workflow providers and LLMs, giving you a unified conversational interface for execution, monitoring, and approval workflows, optimized for production reliability and scale

Chat-First Automation Platform

Conversational control center for workflows, agents, and automations

📋 Overview

AutoPilot is a self-hosted, chat-first automation platform. Instead of manual triggers and complex dashboards, you interact with your automations through natural language. It connects to multiple workflow providers (n8n, Make, Zapier, custom webhooks) and LLM providers, giving you a unified conversational interface for execution, monitoring, file processing, and approval workflows.

✨ Features

AI-Powered Chat

Natural language workflow control with real-time SSE streaming
Mastra-based agent runtime with deterministic orchestrator fallback
Multi-step ReAct agent loop with tool calls (workflow execution, approvals, context retrieval)
Context-aware memory across threads with pgvector-backed retrieval
MCP (Model Context Protocol) support for external tool integrations

Auto Model Router

Automatically discovers available models across all configured LLM providers
Scores and ranks candidates by quality, latency, and health stats
Circuit breaker per provider — failed providers cool down before retries
Prefers reasoning-capable models for complex turns

Workflow Providers

Provider	Type
n8n	Self-hosted / cloud
Make (Integromat)	Cloud scenarios
Zapier	Webhook Zaps
Custom HTTP	Any webhook endpoint
Simulator	Local testing (no external dependency)

LLM Providers

Provider	Notes
OpenAI	GPT-4o, GPT-4o-mini, and any compatible model
Google Gemini	Gemini 2.5 Flash and others
Ollama	Self-hosted Llama, Mistral, etc.
Groq	Fast inference
Mistral	European models + Mistral OCR for PDF/image extraction
OpenAI-Compatible	Any provider with an OpenAI-compatible API

Embeddings & Semantic Search

pgvector in PostgreSQL — no separate vector DB required
Indexes: workflows, workflow runs, document chunks
Embedding providers: Gemini, BGE-small (local), all-MiniLM-L6-v2 (local)
RAG: retrieval-augmented responses grounded in uploaded documents
Semantic search APIs for workflows, runs, and attachments
Similar workflow and related run recommendations

File Attachments

Upload documents (PDF, Word, Excel, PowerPoint, images, audio, text, JSON, XML)
Automatic text extraction, chunking, and embedding
Provider-agnostic OCR chain: configured LLM → Mistral OCR → local fallback
Virus scanning: ClamAV (daemon) or HTTP-based scanner
Up to 25 MB per file, 6 files per message (configurable)

Approvals

Approval-gated workflow execution paths
In-app approval queue with in-browser and web push notifications
Auto-approval mode configurable per deployment

Notifications

Server-Sent Events (SSE) for real-time in-app updates
Web Push (VAPID) for browser notifications
Per-thread notification tracking

Authentication & Security

Email/password + Google OAuth sign-in
TOTP-based MFA (optional per user)
AES-256-GCM encrypted API key storage
CSRF protection, rate limiting, security headers
Webhook HMAC signature verification
Structured request tracing middleware

Observability

Prometheus-compatible metrics via /health/metrics (auth-gated)
Prometheus Pushgateway support
Runtime telemetry for ReAct agent loop steps
Health and readiness endpoints: /health, /health/ready

🚀 Quick Start

Prerequisites

Docker
PostgreSQL 16+ (or use Docker Compose below)

Docker Run

# Pull the image
docker pull ajayduddi/autopilot:latest

# Copy and configure environment variables
cp .env.example .env
# Edit .env with your actual values (database, secrets, callbacks)

# Run with environment file
docker run -d \
  --name autopilot \
  -p 3000:3000 \
  --env-file .env \
  ajayduddi/autopilot:latest

Docker Compose

Create a docker-compose.yml file:

services:
  autopilot:
    image: ajayduddi/autopilot:latest
    restart: always
    env_file: .env.docker
    environment:
      - DATABASE_URL=postgresql://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-autopilot}
    depends_on: [db]
    command: >
      sh -c "bun /app/apps/backend/dist/db/migrate.js &&
             bun /app/apps/backend/dist/index.js"
    ports:
      - "3000:3000"
    volumes:
      - autopilot_data:/data

  db:
    image: pgvector/pgvector:pg16
    environment:
      POSTGRES_USER: postgres
      POSTGRES_PASSWORD: postgres
      POSTGRES_DB: autopilot
    volumes:
      - postgres_data:/var/lib/postgresql/data

volumes:
  autopilot_data:
  postgres_data:

Then run:

# Create compose-specific env file
cp .env.example .env.docker

# Keep DB credentials in one place for both services
echo "POSTGRES_USER=postgres" >> .env.docker
echo "POSTGRES_PASSWORD=postgres" >> .env.docker
echo "POSTGRES_DB=autopilot" >> .env.docker

# Remove DATABASE_URL from .env.docker if present;
# compose builds it from POSTGRES_* and service hostname 'db'.

# Start services
docker compose up -d

Note: Use pgvector/pgvector:pg16 (not plain postgres:16) so the vector extension is available for semantic search and RAG.

⚙️ Environment Variables

Required

Variable	Description
`DATABASE_URL`	PostgreSQL connection string
`AUTH_COOKIE_SECRET`	Session cookie signing secret (32+ chars)
`PROVIDER_API_KEY_ENCRYPTION_KEY`	AES-256-GCM key for stored API keys (32+ chars)
`WEBHOOK_CALLBACK_SECRET`	Fallback HMAC secret for webhook callbacks (32+ chars)
`CALLBACK_BASE_URL`	Public HTTPS URL for receiving workflow callbacks
`FRONTEND_ORIGIN`	Public HTTPS URL for CORS and OAuth redirects
`VITE_API_URL`	Backend API base URL used by frontend at build time

🔐 Authentication

Variable	Default	Description
`GOOGLE_CLIENT_ID`	—	Google OAuth client ID
`GOOGLE_CLIENT_SECRET`	—	Google OAuth client secret
`GOOGLE_REDIRECT_URI`	—	Must be `<FRONTEND_ORIGIN>/api/auth/google/callback`
`SESSION_TTL_DAYS`	`14`	Session lifetime in days
`TRUST_PROXY`	—	Set to `1` when behind a reverse proxy

🔔 Push Notifications (optional)

# Generate VAPID keys once:
npx web-push generate-vapid-keys

VAPID_PUBLIC_KEY=...
VAPID_PRIVATE_KEY=...
VAPID_SUBJECT=mailto:you@example.com

🛡️ File Scanning (optional)

Variable	Description
`CLAMAV_HOST`	ClamAV daemon host
`CLAMAV_PORT`	ClamAV daemon port (default `3310`)
`ATTACHMENT_SCAN_HTTP_URL`	Alternative HTTP-based scanner endpoint
`ATTACHMENT_SCAN_HTTP_TOKEN`	Bearer token for HTTP scanner

🧪 Stability Rollout Flags (Compatibility-First)

Flag	Default	What it controls
`FEATURE_STRUCTURED_LOGGING`	`false`	Enables structured logging format for operational observability pipelines.
`FEATURE_CROSS_ORIGIN_ISOLATION`	`false`	Enables cross-origin isolation headers/behavior for advanced browser runtime features.

🔎 Embeddings & RAG (optional)

Variable	Default	Description
`EMBEDDING_PROVIDER`	`api`	`api`, `bge_local`, or `minilm_local`
`EMBEDDING_MODEL`	`text-embedding-004`	Model ID for the selected provider
`SEMANTIC_SEARCH_TOP_K_DEFAULT`	`8`	Results returned per semantic query
`SEMANTIC_SEARCH_MIN_SCORE`	`0.65`	Minimum cosine similarity threshold
`RAG_MAX_CHUNKS`	`8`	Max chunks injected into assistant context
`RAG_CHUNK_TOKEN_BUDGET`	`2400`	Token budget for grounded context

Local embeddings (no Gemini API key needed):

EMBEDDING_PROVIDER=bge_local
EMBEDDING_MODEL=Xenova/bge-small-en-v1.5

# MiniLM
EMBEDDING_PROVIDER=minilm_local
EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2

🧭 Runtime Configuration

Advanced settings live in ~/.autopilot/config.json (path set by AUTOPILOT_HOME). Key options:

{
  "uploadDir": "./uploads",
  "approvalMode": "default",
  "OLLAMA_URL": "http://ollama.internal:11434",
  "AGENT_RUNTIME_MAX_STEPS": 6,
  "AGENT_MCP_ENABLED": false,
  "AGENT_MCP_SERVERS_JSON": {},
  "MAX_UPLOAD_MB": 25,
  "MAX_FILES_PER_MESSAGE": 6,
  "CONTEXT_MODE_ENABLED": true,
  "CONTEXT_MODE_TARGET_WINDOW_TOKENS": 250000,
  "CONTEXT_MODE_HISTORY_BUDGET_TOKENS": 160000,
  "CONTEXT_MODE_TTL_DAYS": 30,
  "LLM_GENERATE_REPLY_TIMEOUT_MS": 30000,
  "AUTO_ROUTER_DISCOVERY_BREAKER_FAILURES": 3,
  "ATTACHMENT_SCAN_MODE": "off",
  "METRICS_ALLOW_PUBLIC": false,
  "METRICS_AUTH_TOKEN": ""
}

/health/metrics is private by default. Set METRICS_AUTH_TOKEN for authenticated Prometheus scraping, or keep it behind private network access.

🏗️ Architecture

┌──────────────────────────────────────────────────────────┐
│                   Browser (SolidJS)                      │
│   Chat · Threads · Workflows · Approvals · Notifications |
│                       Settings                           │
└───────────────────────┬──────────────────────────────────┘
                        │ REST + SSE
                ┌───────▼────────┐
                │  Express API   │
                │   (Bun)        │
                └───────┬────────┘
         ┌──────────────┼──────────────────┐
         │              │                  │
  ┌──────▼──────┐ ┌─────▼──────┐  ┌───────▼───────┐
  │ PostgreSQL  │ │  Workflow  │  │  LLM Provider │
  │ + pgvector  │ │  Adapters  │  │  (Auto-Router)│
  │  (Drizzle)  │ │ n8n/Make/  │  │ OpenAI/Gemini │
  └─────────────┘ │ Zapier/... │  │ Ollama/Groq/..│
                  └─────┬──────┘  └───────────────┘
                        │ callbacks
              ┌─────────▼──────────┐
              │  /api/webhooks/*   │
              └────────────────────┘

Runtime Model

Every chat message goes through a two-tier runtime:

AgentService (primary) — Mastra-based ReAct agent with tool calls: workflow execution, approval handling, context/memory retrieval, system queries
OrchestratorService (fallback) — deterministic intent-parse + reply-generate path when the agent runtime fails or times out

Streaming uses SSE. Non-streaming uses regular JSON responses. Both paths produce identical structured block output.

🛠️ Development

🧑‍💻 Local Setup

git clone https://github.com/Ajayduddi/AutoPilot.git && cd autopilot
bun install

# Environment
cp .env.example .env
# Edit .env with your credentials

# Database setup
bun --filter=backend run db:preflight
bun --filter=backend run db:migrate

# Start everything
bun run dev

Default URLs:

Frontend: http://localhost:5173
Backend API: http://localhost:3000
Health: http://localhost:3000/health

⌨️ Commands

bun run dev           # Start frontend + backend in watch mode
bun run build         # Build all workspaces
bun run typecheck     # TypeScript check across all workspaces
bun run test          # Run all tests
bun run lint          # Lint backend

# Backend DB workflow
bun --filter=backend run db:preflight     # Validate DB connectivity + extension
bun --filter=backend run db:generate      # Generate migration from schema diff
bun --filter=backend run db:migrate       # Apply pending migrations
bun --filter=backend run db:seed          # Seed development data
bun --filter=backend run db:repair        # Analyze + repair migration journal

🗄️ Database Safety

Production schema changes must follow this order:

db:preflight — verify DB connection and pgvector extension
db:repair:analyze / db:repair:apply — fix journal if needed
db:generate — generate migration SQL from Drizzle schema diff
db:migrate — apply to target database

db:push is blocked in production (use db:push:guarded only for local dev).

📁 Project Structure

apps/
  backend/src/
    config/          # Runtime config loading, schema, persistence
    db/              # Drizzle schema, migrations, seed, preflight, repair
    middleware/       # auth, csrf, rate-limit, security-headers, webhooks, tracing
    providers/
      llm/           # OpenAI, Gemini, Ollama providers + LLMFactory
      workflow/      # n8n, Make, Zapier, Custom, Sim adapters
      embedding/     # Gemini, BGE-local, MiniLM, OpenAI-compatible
    routes/          # Express route handlers
    repositories/    # Data access layer (Drizzle queries)
    services/
      agent/         # Mastra ReAct agent orchestration
      agent-runtime/ # MCP, tools (workflow/approval/context/system)
      ai-routing/    # Auto model router with circuit breaker
      orchestrator/  # Deterministic fallback orchestrator
      context/       # Thread context window assembly
      retrieval/     # Embeddings, RAG, semantic search, recommendations
      attachments/   # Upload processing, OCR, scanning
      notifications/ # SSE events, web push
      auth/          # Session management, Google OAuth, MFA
      settings/      # Provider config management
      extraction/    # Document parsing and chunking
      telemetry/     # ReAct step telemetry
      workflow/      # Workflow CRUD and run management
  frontend/src/
    routes/          # Chat, threads, workflows, approvals, notifications, settings
    components/      # chat/, layout/, notifications/, settings/, workflow/, ui/
    context/         # Auth, notifications, mobile menu, theme contexts
    lib/             # API client, utilities
packages/
  shared/src/        # Shared TypeScript types: chat blocks, API contracts, workflows

🔌 Webhook Integration

AutoPilot receives completion callbacks from workflow providers at:

POST /api/webhooks/callback

All endpoints verify x-webhook-secret HMAC. Secrets are managed per-callback-key from Settings → Webhook Callbacks.

Universal callback format:

{
  "traceId": "trace_abc123",
  "workflowKey": "daily_report",
  "provider": "n8n",
  "status": "completed",
  "result": { }
}

🩺 Troubleshooting

Container won't start

# Check logs
docker logs autopilot

Check DATABASE_URL, AUTH_COOKIE_SECRET, and PROVIDER_API_KEY_ENCRYPTION_KEY are all set.

pgvector extension missing Use pgvector/pgvector:pg16 image (not plain postgres:16), or run CREATE EXTENSION vector; manually, then re-run db:preflight.

Webhook callbacks not arriving

Verify CALLBACK_BASE_URL is a publicly reachable HTTPS URL
Check webhook secret matches the configured key in Settings
Confirm ingress/firewall allows POST /api/webhooks/*

LLM errors

Verify API key in Settings → AI Providers
For Ollama: ensure OLLAMA_URL points to a running Ollama instance accessible from the container
Check model name is correct for the configured provider

Embeddings not working

EMBEDDING_PROVIDER=gemini requires a Gemini API key configured in Settings
For local providers, first run may download model weights (~100 MB)
After changing providers, reindex embeddings from Settings → Retrieval

📚 Documentation

Architecture — runtime model, component map, request flows
Deployment Guide — production deployment details
API Reference — endpoint documentation
Migrations — database migration workflow
DB Migration Safety — migration safety rules
Model Policy — LLM provider and model policies
n8n Contract — n8n webhook integration contract

📄 License

Apache License 2.0 - See LICENSE for details.

Name		Name	Last commit message	Last commit date
Latest commit History 20 Commits
.github/workflows		.github/workflows
apps		apps
docs		docs
packages/shared		packages/shared
scripts		scripts
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
AGENTS.md		AGENTS.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
README.svg		README.svg
bun.lock		bun.lock
bunfig.toml		bunfig.toml
package.json		package.json
tsconfig.base.json		tsconfig.base.json

Folders and files

Latest commit

History

Repository files navigation

📋 Overview

✨ Features

AI-Powered Chat

Auto Model Router

Workflow Providers

LLM Providers

Embeddings & Semantic Search

File Attachments

Approvals

Notifications

Authentication & Security

Observability

🚀 Quick Start

Prerequisites

Docker Run

Docker Compose

⚙️ Environment Variables

Required

🔐 Authentication

🔔 Push Notifications (optional)

🛡️ File Scanning (optional)

🧪 Stability Rollout Flags (Compatibility-First)

🔎 Embeddings & RAG (optional)

🧭 Runtime Configuration

🏗️ Architecture

Runtime Model

🛠️ Development

🧑‍💻 Local Setup

⌨️ Commands

🗄️ Database Safety

📁 Project Structure

🔌 Webhook Integration

🩺 Troubleshooting

📚 Documentation

📄 License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

Packages