Self-hosted multi-agent AI orchestration — structured by design, auditable by default.

swarmwright.com  ·  Docs  ·  Managed hosting  ·  CC BY-NC 4.0

Build teams of AI agents that handle real work. SwarmWright enforces a strict topology — every connection is declared, every action is logged, anything that needs a human lands in the Inbox.

Agents don't call each other freely. You define who talks to whom, what triggers a run, and what requires sign-off. The runtime enforces it.

docker pull ralphbarendse/swarmwright:latest

docker run -d \
  --name swarmwright \
  --network host \
  --restart unless-stopped \
  -v ./data:/data \
  ralphbarendse/swarmwright:latest

Open http://localhost:5001. On first visit you'll be taken to a setup page to create your admin account. After that, go to Settings and enter your LLM provider and API key.

Don't want to run servers? A managed version is available at console.swarmwright.com — the same product on a dedicated, isolated instance with its own subdomain and encryption key. See pricing.

You build swarms — groups of agents, each with a written constitution describing their role and behaviour. Swarms live inside workspaces (think: departments). A company-wide layer sits above everything and is shared across all workspaces.

A run starts when an event fires — on a schedule, via webhook, or manually from the Control Room. The runtime walks the declared topology, calls each agent in turn, logs every step, and surfaces anything requiring human judgement to the Inbox.

Every agent has one of four roles — the topology only allows edges that make architectural sense:

• Canvas — drag-and-drop topology editor, wire agents and set edge types
• Constitution editor — write agent roles in plain Markdown with live preview
• Control Room — fire runs manually, monitor active runs, stop them mid-flight
• Inbox — human-in-the-loop approvals and escalations
• Library — manage knowledge docs, Python skills, and triggers at any scope
• Settings — LLM provider, model, API keys (encrypted at rest), custom branding
• Users — create and manage user accounts, assign per-user permissions

• Topology-enforced agent graphs — delegate, escalate, and report edges only where declared
• Three-scope resource resolver: swarm → workspace → company, unqualified refs auto-resolve
• Python skills — drop a .py file in skills/, agents can call it as a tool
• Scheduled and webhook triggers with cron expressions, backed by YAML files
• Full run trace with per-step input/output logging
• Human-in-the-loop escalation with inbox approvals
• Encrypted secret storage (Fernet, key auto-generated on first boot)
• File-backed topology — hierarchy.json and constitutions are plain files you can version
• Single Docker container, SQLite, no external dependencies
• Built-in user authentication — session-based login, bcrypt passwords, first-boot setup wizard
• Per-user permission system — 15 granular flags covering workspaces, swarms, runs, inbox, skills, and admin

Everything that survives a container rebuild lives in the mounted /data directory:

data/
├── swarm.db                  — SQLite database
├── .encryption_key           — Fernet master key (back this up)
├── company/                  — Company-wide knowledge, skills, triggers
└── workspaces/
    └── <workspace>/
        ├── knowledge/
        ├── skills/
        ├── triggers/         — YAML trigger definitions
        └── swarms/
            └── <swarm>/
                ├── meta.yaml
                ├── hierarchy.json    — topology (source of truth)
                └── agents/           — constitution .md files per agent

The GUI writes hierarchy.json; you can also edit it directly and the runtime picks up changes live.

All required settings (LLM provider, API keys) can be configured through the Settings UI. Environment variables are optional overrides — useful for automated deployments or secret managers.

CC BY-NC 4.0 — free for personal and non-commercial use.