Autosclaw

Autonomous AI Agent Orchestration Platform

Spawn, schedule, and orchestrate autonomous Claude-powered agents — each in its own Docker container — from a real-time web dashboard.

---

Table of Contents

• What is autosclaw?
• Core Functionality
• Architecture
• Getting Started (AI-Assisted Setup)
• Configuration
• How It Works
• Security
• API Reference
• Tech Stack
• Project Structure
• Contributing
• License

What is autosclaw?

autosclaw is a self-hosted platform that turns Claude into an army of autonomous agents. Each agent runs in its own isolated Docker container with full tool access — they can write code, browse the web, manage files, interact with GitHub, and even spawn other agents to delegate work.

From the real-time web dashboard, you watch agents think, act, and collaborate. You see every tool invocation, every token spent, every dollar of cost — live. Schedule agents to run on cron expressions for recurring tasks like daily reports, repository maintenance, or automated code reviews. Or spawn them on-demand and let them loose.

What makes autosclaw unique is its multi-agent orchestration: agents aren't isolated workers — they're a network. Any agent can spawn, list, and manage other agents through MCP tools, enabling complex multi-step workflows where agents coordinate, delegate, and build on each other's work. All while you watch from a single pane of glass.

Core Functionality

Autonomous AI Agents

Each agent gets its own Docker container running Claude Opus 4.6 with the full Claude Code toolset:

• Code execution — Bash, file read/write/edit, glob, grep
• Web access — web search, web fetch, browsing
• Development tools — Git, GitHub CLI (gh), plan mode, worktrees
• Agent orchestration — spawn sub-agents, manage schedules, delegate work
• Human-in-the-loop — agents can pause and ask you questions mid-task, then continue with your answer

Multi-Agent Orchestration via MCP

Agents form an interconnected network through 10 MCP (Model Context Protocol) tools:

Tool	Description
spawn_agent	Spawn a child agent with inherited permissions
list_agents	List all running agents (with isSelf flag)
despawn_agent	Stop an agent (cannot despawn self)
create_schedule	Create a cron-based recurring or one-time schedule
list_schedules	List all schedules
get_schedule	Get details of a specific schedule
update_schedule	Modify an existing schedule
pause_schedule	Pause an active schedule
resume_schedule	Resume a paused schedule
delete_schedule	Soft-delete a schedule

Key design decisions:

• Recursive spawning — agents spawn children, which spawn grandchildren, and so on
• Permission inheritance — child agents automatically inherit their parent's permissions
• Safety guardrails — agents cannot despawn themselves or agents already stopping

Real-Time Dashboard

A React 19 progressive web app that streams everything live over WebSocket:

• Live output — watch agent text, collapsible thinking blocks, and system messages as they happen
• Tool cards — structured display of every tool invocation with syntax-highlighted inputs and diffs
• Interactive chat — jump into any running agent's session at any time to provide guidance, course-correct, or give additional instructions mid-task
• Ask-user flow — agents can pause for user input; you get a push notification and see selectable options in the dashboard
• Sortable tables — sort agents and schedules by any column (name, status, tokens, cost)
• Bulk actions — select multiple agents and stop them in one click
• Agent history — browse output and tool logs for stopped or deleted agents
• PWA support — installable on mobile/desktop, push notifications, offline-ready service worker

Cron Scheduling

Automate recurring work with cron-powered scheduling:

• Recurring schedules — standard 5-field cron expressions (UTC timezone)
• One-time schedules — fire once and auto-delete
• Pause / resume — temporarily disable schedules without deleting them
• Custom agent config — each schedule defines the spawned agent's name, purpose, and system prompt
• Permission scoping — schedules inherit the creator's permissions
• 30-second tick — the scheduler checks and fires every 30 seconds

Cost & Token Tracking

Full visibility into what your agents are spending:

• Per-agent breakdown — input tokens, output tokens, cache read, cache creation, and USD cost
• Live updates — token counters update in real-time as agents work
• Usage dashboard — paginated table of all agents sorted by cost, with totals
• Summary cards — at-a-glance totals for running agents, total agents, and schedules

Role-Based Access Control

Three roles with granular permissions:

Role	Capabilities
Admin	Full control: manage all agents, schedules, users, reload service
Member	Create and manage your own agents and schedules
Viewer	Read-only access to the dashboard

Agents also receive granular permissions (agent:spawn, agent:schedule) that propagate down the hierarchy to child agents.

Architecture

autosclaw consists of three main components: the Manager (host process), Agent containers (Docker), and the Dashboard (React UI). They communicate over WebSocket for real-time streaming and REST for control operations.

┌──────────────────────────────────────────────────────────────────────┐
│                      Web Dashboard (React 19 + Vite)                 │
│                                                                      │
│    Live output streaming    Sortable agent/schedule tables           │
│    Tool invocation cards    Token & cost tracking                    │
│    Interactive chat         Push notifications (PWA)                 │
│    Role-based UI            Ask-user question flow                   │
└─────────────────────────────┬────────────────────────────────────────┘
                              │ WebSocket (/ws) — JWT authenticated
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│                        MANAGER  (Host)                               │
│                                                                      │
│  ┌────────────┐  ┌─────────────┐  ┌─────────────┐  ┌────────────┐  │
│  │  Express 5  │  │  WebSocket  │  │  SQLite DB  │  │    Cron    │  │
│  │  REST API   │  │   Servers   │  │  (WAL mode) │  │  Scheduler │  │
│  │             │  │             │  │             │  │             │  │
│  │ /api/spawn  │  │ /ws   (UI)  │  │ • agents    │  │ 30s tick   │  │
│  │ /api/agents │  │ /agent      │  │ • messages  │  │ recurring  │  │
│  │ /api/despawn│  │  (agents)   │  │ • users     │  │ one-time   │  │
│  │ /api/sched. │  │             │  │ • schedules │  │ auto-spawn │  │
│  │ /api/push   │  │ Broadcast   │  │ • config    │  │            │  │
│  │ /api/health │  │ to watchers │  │ • push_subs │  │            │  │
│  └────────────┘  └──────┬──────┘  └─────────────┘  └────────────┘  │
│                          │                                           │
│  SECURITY LAYER          │  Docker CLI: docker run / stop / rm      │
│  ├── JWT auth (bcrypt + 24h tokens)                                  │
│  ├── Per-agent hex tokens (timing-safe comparison)                   │
│  ├── RBAC (admin / member / viewer)                                  │
│  ├── Secrets via stdin (never on disk or in env vars)                │
│  └── Post-startup secret scrubbing from process.env                  │
│                                                                      │
│  RESILIENCE                                                          │
│  ├── 30s health check reconciliation loop                            │
│  ├── Orphan container detection and cleanup                          │
│  └── Graceful shutdown (15s timeout → force exit)                    │
│                                                                      │
└──┬───────────────────────────────────────────────────────────────────┘
   │
   │  WebSocket (/agent) + secrets via stdin
   ▼
┌──────────────────────────────────────────────────────────────────────┐
│                       DOCKER CONTAINERS                              │
│                                                                      │
│  ┌──────────────────┐  ┌──────────────────┐  ┌──────────────────┐   │
│  │   Agent #1        │  │   Agent #2        │  │   Agent #N       │   │
│  │                   │  │                   │  │                  │   │
│  │  Claude Agent SDK │  │  Claude Agent SDK │  │  Claude Agent    │   │
│  │  (Opus 4.6)       │  │  (Opus 4.6)       │  │  SDK (Opus 4.6) │   │
│  │                   │  │                   │  │                  │   │
│  │  Built-in tools:  │  │  MCP tools:       │  │  Permissions     │   │
│  │  Bash, Read,      │  │  spawn_agent      │  │  inherited from  │   │
│  │  Write, Edit,     │  │  list_agents      │  │  parent          │   │
│  │  Grep, Glob,      │  │  despawn_agent    │  │                  │   │
│  │  WebSearch,       │  │  schedule CRUD    │  │  git, gh CLI     │   │
│  │  WebFetch, Agent  │  │  (7 tools)        │  │                  │   │
│  └──────────────────┘  └──────────────────┘  └──────────────────┘   │
│           ▲                     ▲                     ▲              │
│           └─────────────────────┴─────────────────────┘              │
│                    Agent-to-Agent Orchestration                       │
└──────────────────────────────────────────────────────────────────────┘

Component Responsibilities

Manager (manager/src/) — the central orchestrator running on the host:

Module	Responsibility
main.ts	Express app init, Docker image build, agent restoration on startup, 30s health check loop, graceful shutdown
api.ts	REST API routes for spawning/despawning agents and schedule CRUD, dual auth (JWT + agent tokens)
auth.ts	User registration/login, JWT generation/verification, password hashing (bcryptjs), role management
communication.ts	Two WebSocket servers — /ws for the dashboard and /agent for agent containers — message routing, permission checks, project management
containers.ts	Docker container lifecycle: image building, agent spawning (with stdin secret delivery), stopping, orphan cleanup
db.ts	SQLite schema and CRUD for agents, messages, schedules, users, projects, and config (WAL mode, soft deletes)
schedules.ts	Cron scheduler: 30-second tick loop, next-run computation, auto-spawn on match, one-time schedule cleanup
push.ts	Web Push notifications via VAPID keys for agent completion, errors, input requests, and schedule triggers
logger.ts	Structured JSON logging with file rotation (10 MB max, 5 files)

Agent (agent/src/) — runs inside each Docker container:

Module	Responsibility
main.ts	Entrypoint: reads secrets from stdin, creates MCP server with 10 tools, runs Claude Agent SDK session, scrubs secrets from process.env
manager-client.ts	WebSocket client to manager: registration, output/token/tool-use streaming, ask-user flow, spawn requests (60s timeout)
types.ts	SDK message type helpers for display formatting

Dashboard (manager/ui/src/) — React 19 single-page app:

Module	Responsibility
context/AppContext.tsx	Global state (agents, schedules, projects, outputs) via React Context + Reducer, WebSocket connection lifecycle
hooks/	useAuth (JWT mgmt), useAutoScroll (smart scrolling), useConfirmDialog, usePushNotifications, useTableSort
components/dashboard/	Main view: SummaryCards, AgentTable, ScheduleTable, AgentUsageTable, ProjectsPanel
components/agent/	Agent detail: OutputArea (text/thinking), ToolCard (syntax-highlighted), ChatBar, TokenBar, AskUserCard
components/modals/	NewAgentModal (spawn form), NewScheduleModal (cron config), ConfirmDialog

Communication Protocol

Agent ↔ Manager (WebSocket /agent):

Direction	Message Type	Purpose
Agent → Manager	register	Authenticate with agent ID and token
Agent → Manager	output	Text output (text, thinking, result, system)
Agent → Manager	tool_use	Tool invocation details
Agent → Manager	tokens	Token usage and cost update
Agent → Manager	spawn_agent	Request to spawn a child agent
Agent → Manager	ask_user	Pause and request user input
Manager → Agent	chat	Message from user via dashboard
Manager → Agent	ask_user_response	User's answer to a pending question
Manager → Agent	close	Shutdown order

Dashboard ↔ Manager (WebSocket /ws): The manager broadcasts agent_list and schedule_list on every state change, and streams output, tool_use, tokens, and ask_user events in real-time to watchers.

Data Model

SQLite (WAL mode) with the following tables:

Table	Key Fields
agents	id, name, token, purpose, system_prompt, cost_usd, input/output/cache tokens, status, permissions, project_id
messages	agent_id, role, msg_type (text/tool_use), text, metadata (JSON)
schedules	id, name, cron_expression, schedule_type (recurring/once), agent config, status, next_run_at
users	id, username, password_hash, role (admin/member/viewer)
projects	id, name, purpose, github_token
config	key/value pairs (VAPID keys, etc.)
push_subscriptions	endpoint, keys_p256dh, keys_auth

All deletions are soft deletes (preserving audit history).

Getting Started (AI-Assisted Setup)

autosclaw includes an AI-assisted setup experience powered by a Claude Code custom skill. Instead of manually configuring environment variables and running build steps, Claude walks you through the entire process interactively — checking prerequisites, generating secrets, and configuring everything for you.

Prerequisites

• Node.js 22+
• Docker (running and accessible)
• GitHub CLI (gh) (install)
• Anthropic API Key (get one here)
• Claude Code (install)

Setup

1. Clone the repository

gh repo clone BreuerFlorian/autosclaw
cd autosclaw

1. Run the AI-assisted setup

Open Claude Code in the project directory and run the setup skill:

/setup

Claude will interactively guide you through:

• Verifying prerequisites (Node.js, npm, Docker)
• Building the agent Docker image
• Configuring your .env file (API key, admin credentials, JWT secret, optional settings)
• Installing manager dependencies
• Building the React dashboard
• Starting the manager
• Optionally installing a systemd service for auto-start on boot

The setup skill handles secret generation, input validation, and error recovery — you just answer the prompts.

Manual Setup (Alternative)

If you prefer to set up manually without AI assistance:

1. Build the agent Docker image:

cd agent && docker build -t autosclaw-agent . && cd ..

1. Create a .env file in the project root:

ANTHROPIC_API_KEY=sk-ant-...
ADMIN_USERNAME=admin
ADMIN_PASSWORD=your-secure-password
JWT_SECRET=your-jwt-secret
GH_TOKEN=ghp_...              # Optional: enables GitHub self-modification

1. Install and run:

cd manager && npm install && cd ui && npm install && npm run build && cd .. && npm start

The dashboard is now running at http://localhost:4000.

Systemd Service (Optional)

For production/server deployments, install as a systemd user service:

./install.sh

This creates autosclaw-manager.service that auto-starts on boot, restarts on failure, and runs in the background. Management commands:

systemctl --user status autosclaw-manager
systemctl --user restart autosclaw-manager
journalctl --user -u autosclaw-manager -f

Configuration

Environment Variables

Variable	Default	Description
ANTHROPIC_API_KEY	required	Your Anthropic API key for Claude
ADMIN_USERNAME	—	Username for the initial admin account (seeded on first run)
ADMIN_PASSWORD	—	Password for the initial admin account
JWT_SECRET	random	Secret for signing JWT tokens. Set this for persistent sessions across restarts
PORT	4000	Port the manager server listens on
MANAGER_HOST	auto-detected	Docker bridge gateway IP. Auto-detected on Linux via ip route
ALLOW_REGISTRATION	false	Set to true to enable user self-registration
LOG_LEVEL	info	Logging verbosity (debug, info, warn, error)
LOG_DIR	—	Directory for log files (with automatic rotation at 10 MB, keeps 5 files)
GH_TOKEN	—	GitHub personal access token. Gives agents the ability to interact with GitHub repositories — clone, push, open PRs, and self-modify the autosclaw codebase itself

Self-Modification with GH_TOKEN

When you provide a GH_TOKEN in your .env, agents gain full GitHub access through the gh CLI. This enables a powerful self-modification capability: agents can clone the autosclaw repository, make changes to its own codebase, commit, push, and open pull requests — effectively allowing the platform to improve itself.

Use cases include:

• Agents that refactor or extend autosclaw's own source code
• Automated code review and PR creation on any repository
• Scheduled agents that maintain repositories (dependency updates, changelog generation)
• Multi-agent workflows where one agent writes code and another reviews it via GitHub

The token is delivered to agent containers securely via stdin (same as all other secrets) and scrubbed from process.env after startup.

How It Works

Agent Lifecycle

1. Spawn — You spawn an agent from the dashboard (or an agent spawns a child via MCP). The Manager generates a UUID and a unique 32-byte hex token, then runs docker run to create a new container from the autosclaw-agent image.

2. Secret delivery — The API key, manager URL, agent token, and optional GitHub token are piped to the container via stdin. They never touch disk or container environment variables. After the SDK subprocess starts, all secrets are scrubbed from process.env.

3. Registration — The agent connects back to the Manager via WebSocket, authenticates with its token, and begins its autonomous Claude session with the configured system prompt.

4. Execution — The agent runs autonomously, executing tools (Bash, file operations, web search, git) and reporting output, tool invocations, and token usage to the Manager in real-time.

5. Orchestration — Through MCP tools, the agent can spawn child agents, create schedules, list running agents, and despawn agents — enabling recursive multi-agent workflows with inherited permissions.

6. Live guidance — You can jump into any running agent's session at any time via the dashboard chat. Send instructions, course-correct, or provide context — the agent incorporates your message and continues. When the agent itself needs input, it sends an ask_user message with selectable options, and you get a push notification.

7. Completion — When the agent finishes (or is stopped), the container exits and is auto-removed (--rm). The Manager detects the exit, updates the database, and broadcasts the state change to the dashboard.

Scheduling

The Manager's cron scheduler ticks every 30 seconds. For each active schedule whose next_run_at has passed, it spawns an agent with the schedule's configured name, purpose, system prompt, and permissions. One-time schedules are auto-deleted after firing.

State Persistence

All state is stored in SQLite (WAL mode). On startup, the Manager reconciles its database with running Docker containers — restoring active agents and cleaning up orphans (containers without a matching DB entry, or DB entries without a running container).

Security

Secret Management

Secrets are piped to containers via stdin only — never stored in environment variables, CLI arguments, Docker labels, or on disk. After the Claude SDK subprocess starts, the agent scrubs ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, GITHUB_TOKEN, AGENT_TOKEN, and others from process.env, preventing leakage through /proc, environment dumps, or child processes.

Authentication

• User auth — username/password login produces a 24-hour JWT. Passwords hashed with bcryptjs (10 rounds).
• Agent auth — each container gets a unique 32-byte hex token verified with crypto.timingSafeEqual (timing-safe comparison).
• Dual auth middleware — every API endpoint accepts either a user JWT or an agent token with appropriate permission checks.

Container Isolation

Every agent runs in a separate Docker container with --rm (auto-cleaned on exit). No host filesystem access, no privileged mode, no shared volumes.

Resilience

• 30-second reconciliation loop — detects and cleans up orphan containers
• Graceful shutdown — stops all agents, closes WebSockets, waits 15 seconds before force-exit
• Structured logging — JSON-formatted logs with request IDs and automatic file rotation
• Soft deletes — agents and schedules are never hard-deleted, preserving audit history

Input Validation

All user inputs are validated at the API boundary: agent names (255 chars), purposes (1,000 chars), system prompts (100,000 chars), chat messages (50,000 chars), and cron expressions (validated with cron-parser).

API Reference

REST API

All /api/* endpoints require Authorization: Bearer <token> (JWT or agent token).

Method	Endpoint	Description
GET	/api/health	Health check (no auth required)
POST	/api/spawn	Spawn a new agent container
POST	/api/despawn	Stop and remove an agent
GET	/api/agents	List all agents
GET	/api/schedules	List all schedules
POST	/api/schedules	Create a new schedule
GET	/api/schedules/:id	Get schedule details
PUT	/api/schedules/:id	Update a schedule
DELETE	/api/schedules/:id	Soft-delete a schedule
POST	/api/schedules/:id/pause	Pause a schedule
POST	/api/schedules/:id/resume	Resume a paused schedule
POST	/api/push/subscribe	Subscribe to push notifications
POST	/api/push/unsubscribe	Unsubscribe from push notifications
GET	/api/push/subscriptions	List push subscriptions

Authentication Endpoints

Method	Endpoint	Description
POST	/auth/login	Authenticate and receive JWT (24h expiry)
POST	/auth/register	Register a new user (when ALLOW_REGISTRATION=true)
GET	/auth/me	Verify the current token

WebSocket Endpoints

Endpoint	Purpose
/ws?token=<jwt>	Dashboard — real-time agent list, output streaming, cost updates
/agent	Agent-to-Manager — registration, message relay, token usage, spawn requests

Tech Stack

Layer	Technology	Purpose
Runtime	Node.js 22 + TypeScript 5.6	Server runtime with type safety
Server	Express 5	HTTP server and REST API
Real-Time	ws (WebSocket)	Bidirectional communication
Database	better-sqlite3 (WAL mode)	Embedded persistent storage
Auth	bcryptjs + jsonwebtoken	Password hashing + JWT tokens
Scheduling	cron-parser	Cron expression parsing and scheduling
Notifications	web-push (VAPID)	Push notifications to browsers and mobile
Logging	Custom JSON logger	Structured logs with file rotation
AI	@anthropic-ai/claude-agent-sdk	Autonomous Claude agent sessions
Tools	@modelcontextprotocol/sdk	MCP server for inter-agent orchestration
Validation	Zod 4	Schema validation for MCP tool inputs
Containers	Docker (node:22-slim)	Isolated agent execution
Frontend	React 19 + Vite	Real-time dashboard UI

Project Structure

autosclaw/
│
├── install.sh                       # systemd user service installer
├── .env                             # Environment variables (created during setup)
│
├── .claude/
│   └── commands/
│       └── setup.md                 # AI-assisted setup skill for Claude Code
│
├── agent/                           # Agent — runs inside Docker containers
│   ├── Dockerfile                   # node:22-slim + git + gh CLI
│   ├── build.sh                     # Build helper: docker build -t autosclaw-agent .
│   ├── Claude.md                    # Agent behavior guidelines
│   ├── package.json                 # Dependencies: agent SDK, MCP SDK, ws, zod
│   ├── tsconfig.json
│   └── src/
│       ├── main.ts                  # Entrypoint: stdin secrets, MCP server, SDK session
│       ├── manager-client.ts        # WebSocket client connecting back to Manager
│       └── types.ts                 # SDK message types and display helpers
│
├── manager/                         # Manager — runs on the host
│   ├── start.sh                     # Full startup script (pull, install, build, run)
│   ├── package.json                 # Dependencies: express, ws, better-sqlite3, etc.
│   ├── tsconfig.json
│   ├── src/
│   │   ├── main.ts                  # Express app, Docker image build, agent restoration
│   │   ├── api.ts                   # REST API routes (/api/spawn, /api/agents, etc.)
│   │   ├── auth.ts                  # JWT authentication (/auth/login, /auth/register)
│   │   ├── communication.ts         # WebSocket servers (dashboard /ws + agent /agent)
│   │   ├── containers.ts            # Docker container lifecycle (spawn, stop, secrets)
│   │   ├── db.ts                    # SQLite schema and CRUD (agents, messages, schedules)
│   │   ├── env.ts                   # Environment variable loading and defaults
│   │   ├── logger.ts                # Structured JSON logging with file rotation
│   │   ├── push.ts                  # Web Push notification service (VAPID)
│   │   ├── pushRoutes.ts            # Push subscription API endpoints
│   │   └── schedules.ts             # Cron schedule engine (30s tick loop)
│   │
│   └── ui/                          # React dashboard (built into src/public/)
│       ├── index.html
│       ├── package.json             # React 19, Vite, TypeScript
│       └── src/
│           ├── App.tsx              # Root layout: sidebar + main panel
│           ├── context/
│           │   └── AppContext.tsx    # Global state (React Context + Reducer + WebSocket)
│           ├── hooks/               # useAuth, useAutoScroll, useConfirmDialog, etc.
│           └── components/
│               ├── auth/            # LoginOverlay
│               ├── layout/          # Sidebar, MobileHeader
│               ├── modals/          # NewAgentModal, NewScheduleModal, ConfirmDialog
│               ├── dashboard/       # Dashboard, AgentTable, ScheduleTable, TokenSummary
│               ├── agent/           # AgentHeader, OutputArea, ToolCard, ChatBar, TokenBar
│               ├── schedule/        # ScheduleDetail
│               ├── settings/        # NotificationSettings
│               ├── notifications/   # Toast, ToastContainer
│               └── pwa/             # InstallPrompt

Contributing

Contributions are welcome! Here's how to get started:

1. Fork the repository (gh repo fork BreuerFlorian/autosclaw --clone)
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push and open a Pull Request (gh pr create --title 'Add amazing feature')

Development Setup

For frontend development with hot reload:

cd manager/ui
npm install
npm run dev

The Vite dev server proxies API requests to http://localhost:4000, so make sure the manager is running.

License

This project is maintained by @BreuerFlorian. Please check the repository for license information.

---

Built with Claude Agent SDK by Anthropic

autosclaw — let your agents do the work.