QUESTPIE Autopilot

AI-native company operating system. Your company is a container. Your employees are agents.

---

Quick Start

One-line install (VPS/server)

curl -fsSL https://raw.githubusercontent.com/questpie/autopilot/main/install.sh | bash

Docker

git clone https://github.com/questpie/autopilot
cd autopilot
cp .env.example .env
docker compose up

Open http://localhost:3000 for the dashboard. The local Caddy proxy keeps browser auth and API calls same-origin.

Authentication: Set your OpenRouter API key in .env (OPENROUTER_API_KEY=sk-or-...) or run autopilot provider set openrouter --api-key <key>. One key gives you access to all models (Anthropic, OpenAI, Google, etc.).

bunx (local dev)

bunx @questpie/autopilot init my-company
cd my-company

# Set your OpenRouter API key
autopilot provider set openrouter --api-key sk-or-...

bunx @questpie/autopilot start

---

What Happens

$ autopilot ask "Build a pricing page with Stripe"

CEO Agent decomposing intent...

  task-050: Scope requirements      → Sam (strategist)
  task-051: Design UI               → Jordan (design)
  task-052: Implement with Stripe   → Max (developer)
  task-053: Write copy & announce   → Morgan (marketing)

Sam is starting now. You'll be notified when approvals are needed.

1. You give intent
2. CEO agent decomposes into tasks
3. Agents execute: strategist → planner → developer → reviewer → devops
4. You approve at gates (merge, deploy, spend)
5. Workflow execution is tracked durably in SQLite
6. Files remain the authored source of truth. Every action is still git-auditable.

---

Workflow Runtime Model

Autopilot treats workflows as the primary execution primitive.

• team/workflows/*.yaml defines the authored workflow structure
• the orchestrator compiles workflows into explicit step contracts
• SQLite stores runtime execution in workflow_runs and step_runs
• Durable Streams stores append-only replay timelines for sessions and workflow events

Agents execute the current step. The app owns validation, transition state, deduplication, and replayability.

---

Your Team, Your Rules

Autopilot ships with 8 default agents across 8 roles — meta, strategist, planner, developer, reviewer, devops, marketing, design. Customize them or create your own in team/agents/*.yaml.

# team/agents/max.yaml
id: max
name: Max
role: developer
description: Implementation, tests, debugging
model: anthropic/claude-sonnet-4
tools: [fs, terminal]
fs_scope:
  read: ["**"]
  write: ["**"]

---

Architecture

Human         CLI · Dashboard · Telegram · Webhooks
  │
Orchestrator  Watcher · Workflows · Spawner · Context · Memory · Cron · Durable Streams
  │
Agents        TanStack AI + OpenRouter · Per-Agent Model Picker · 7 Tools · Sandboxed FS · Memory
  │
Storage       SQLite + Drizzle · YAML/MD/JSON · FTS5 + sqlite-vec · Git

• Config is files — YAML, Markdown, JSON. ls your company config.
• Runtime is SQLite — tasks, messages, sessions, workflow runs, step runs, search. Zero external deps.
• Git is the audit trail — every agent action = commit.
• MCP server — expose Autopilot to Claude Desktop/Code via tasks, agents, search, sessions.
• Durable Streams — append-only session and workflow timelines with live tailing and replay.
• Per-agent model picker — assign different models to different agents via OpenRouter.
• One Bun process — orchestrator + API + dashboard. ~100MB RAM.
• Same-origin browser auth — dashboard traffic stays on one origin; API proxying is handled by Vite in dev and Caddy in Docker/self-hosted setups.

---

Self-Hosted

Autopilot is designed to run on YOUR infrastructure. No SaaS lock-in.

Setup	Command	Cost
Local (macOS/Linux)	bunx @questpie/autopilot start	Free (+ API costs)
Docker	docker compose up	Free (+ API costs)
Hetzner VPS	One-click deploy →	€4.35/mo + API
Any VPS	curl ... install.sh | bash	Your VPS + API

→ Full VPS deployment guide

---

CLI Reference

autopilot init <name>        # Create a new company
autopilot start              # Start orchestrator + dashboard
autopilot ask "<intent>"     # Send intent to CEO
autopilot status             # Company overview
autopilot tasks              # List tasks
autopilot agents             # List agents
autopilot attach <agent>     # Watch agent work (live)
autopilot inbox              # Pending approvals
autopilot approve <id>       # Approve action
autopilot reject <id>        # Reject action
autopilot chat <agent>       # Direct chat with agent
autopilot dashboard          # Open web dashboard
autopilot secrets            # Manage API keys
autopilot auth               # Manage authentication
autopilot provider set <p>   # Configure AI provider (openrouter)

---

Company Structure

my-company/
├── company.yaml              # Company configuration
├── team/
│   ├── agents/               # AI agent definitions (one file per agent)
│   ├── humans/               # Human team members (one file per human)
│   ├── roles.yaml            # RBAC role definitions
│   ├── schedules/            # Cron-triggered jobs (one file per schedule)
│   ├── webhooks/             # External webhook integrations (one file per webhook)
│   ├── workflows/            # development (12), marketing (7), incident (8)
│   └── policies/             # Human approval requirements
├── knowledge/                # Searchable knowledge base (markdown)
├── skills/                   # Agent skills (agentskills.io, 20 built-in)
├── dashboard/                # Living dashboard (pins, widgets, pages)
├── tasks/                    # Task status directories (runtime)
├── comms/                    # Communication channels (runtime)
├── logs/                     # Activity & session logs (runtime)
├── context/memory/           # Per-agent persistent memory (runtime)
├── secrets/                  # Encrypted secrets (runtime)
├── projects/                 # Project workspaces (runtime)
├── artifacts/                # Generated apps & content (runtime)
└── .data/autopilot.db        # SQLite (tasks, messages, workflow_runs, step_runs, FTS5, embeddings)

Note: Tasks, messages, activity, workflow runs, and step runs are stored in SQLite (.data/autopilot.db), not as YAML files. Runtime directories are created by autopilot init, not from the template.

Config format: Legacy monolithic files (team/agents.yaml, team/humans.yaml, team/webhooks.yaml, team/schedules.yaml) are no longer supported. Use folder-based config only.

---

Deploy to Hetzner

Best price/performance for self-hosting. ARM64 servers from €4.35/mo.

# On your Hetzner VPS (Ubuntu 24.04):
curl -fsSL https://get.docker.com | sh
mkdir -p /opt/autopilot && cd /opt/autopilot
curl -fsSL https://raw.githubusercontent.com/questpie/autopilot/main/install.sh | bash

Get €20 free credits →

---

Documentation

• Getting Started
• Architecture
• Agents & Roles
• Internal Workflow Memo
• Config Folder Migration
• CLI Reference
• VPS Deployment
• Docker Guide
• Security & Auth

---

Contributing

git clone https://github.com/questpie/autopilot.git
cd autopilot && bun install
bunx turbo build
bunx turbo test

---

License

MIT — QUESTPIE s.r.o. 2026

Built by QUESTPIE