SwarmWright: Structured Multi-Agent AI Defined in Markdowns

SwarmWright — Build the Swarm

SWARM WRIGHT

multi-agent AI, built by you.

A self-hosted workbench for designing, running, and inspecting swarms of AI agents. Agents are plain markdown, orchestrated by one config, observable in real time.

QUICK START — two commands ↓

$ docker pull ralphbarendse/swarmwright:latest

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

free

source-available · CC BY-NC 4.0

self-hosted

read the docs →

SwarmWright — Self-hosted multi-agent AI orchestration platform for agentic LLM workflows

Most AI agent tools pull you to one extreme.

Either you hand over full control and hope for the best, or you hand-code every step yourself. We built SwarmWright for the space in between — structured enough to trust, flexible enough to actually use.

← give away control

SwarmWright lives here

hand-roll everything →

not this

The autonomy trap

Full Autonomy

Autonomous agents · no guardrails · hope for the best

Give agents a goal and let them figure out the rest. They call anything, spawn sub-agents, take actions you never explicitly authorized. Looks impressive in demos. It's a liability in production.

Agents call anything — no enforcement

You can't explain what happened

Important calls happen without review

Your compliance team will never sign off

this

Structured Agent Work

SwarmWright

Agents operate inside a topology you design. Every connection is declared upfront. Every decision leaves a reason. You approve the ones that matter. Non-technical people can read the configuration and know exactly what the system does.

All connections declared — nothing unchecked

Every edge has a stated purpose

Human gates built directly into workflows

Readable by business, ops, and compliance

not this

The rigidity trap

Custom Pipelines

Custom code · DIY pipelines · hard-coded chains

Write every step by hand. It works — until it doesn't. Changing one thing breaks three others. Six months in, nobody understands it. Every new workflow is another one-off only one person can maintain.

Every swarm is a unique implementation

No standard structure to onboard into

Expensive to change, brittle to extend

Auditing means reading source code

SwarmWright is the operating system for your AI workforce — structured enough to deploy with confidence, expressive enough to handle real-world complexity, and every decision is traceable back to the rule that caused it.

Features

opinionated by design. every choice has a reason.

SwarmWright is opinionated by design. Every architectural choice has a reason. Every agent connection must be declared. Every action leaves a trace.

No server to configure, no infrastructure to manage. Pull the container, point it at your API key, and your swarm is live in under a minute. Everything your agents produce is saved locally — your data never leaves your machine.

Works on any machine that runs Docker

All data stored locally in a single folder you own

Survives restarts — pick up exactly where you left off

Your LLM credentials are encrypted and never leave the container

Claude (Anthropic) GPT-4 (OpenAI) Runs on Docker No cloud lock-in Local-first data

Every connection between agents is declared upfront — no surprises, no agents calling things they shouldn't. If an agent tries to do something it wasn't given permission to do, it's stopped and flagged immediately. You stay in control.

Define which agents can talk to which — visually

Each connection has a stated reason, written in plain language

Unauthorized actions are caught and surfaced, not silently skipped

Your compliance team can read the rules — no code required

Drag agents onto a canvas, draw connections between them, and describe why each connection exists. It looks and feels like sketching on paper — because it should. No YAML to write by hand. No diagram that's out of date the moment you deploy.

Drag-and-drop canvas — place agents, draw connections

The diagram is the configuration — always in sync

Each connection asks: what is this for? Forces clear thinking

Color-coded by role so the structure is obvious at a glance

SwarmWright organizes agents the same way good teams are organized — strategy at the top, execution at the bottom, coordination in the middle. Everyone knows who's responsible for what.

■ Policy — sets the rules and goals

■ Orchestrator — coordinates the work

■ Executioner — gets things done

■ Perceptionist — reads and interprets data

Each agent has a constitution — a plain-language document that describes its role, values, and what it knows. Your Finance manager can read it. Your compliance officer can sign off on it. No prompt engineering required.

Written in plain English — readable by anyone

Built-in editor with live preview

Attach company knowledge and procedures directly to agents

Every action an agent takes is recorded with a human-readable reason. No black boxes. You can trace any outcome back to the exact decision that caused it — useful for compliance, debugging, and just knowing what happened.

Every step logged with a plain-language reason

Full history of what each agent did and why

Unexpected behavior flagged clearly, not buried in logs

Agents can run Python scripts, query databases, call external APIs, and process files. Each skill runs as a separate subprocess with a configurable timeout. You decide what tools are available and to whom.

Write custom tools in Python and attach them to agents

Each tool runs as a subprocess with a declared timeout

Tools are reusable across the whole organization

Permissions are explicit — no tool runs without being declared

Agents know when a decision is too important to make alone. They pause and send it to your inbox. You approve, reject, or add context — and the work continues from exactly where it stopped. You're never out of the loop.

Agents ask before acting on high-stakes decisions

Inbox for approvals with full context attached

Work resumes from the exact point it paused

Bring your own Anthropic, OpenAI, or Deepseek key. Switch providers without rebuilding anything. Your credentials are encrypted and stored locally — never sent to a third party. Different swarms can run on different models.

Works with Claude (Anthropic), GPT-4 (OpenAI), and Deepseek

Switch models from the Settings screen — no code changes

API keys encrypted and stored locally

Upload documents, procedures, and reference material once — and make them available to any agent that needs them. Company-wide policies, department rules, or workflow-specific context: everything is organized and stays up to date.

Company-wide documents available to all agents

Department-level rules scoped to the right teams

Workflow-specific context kept private to one swarm

A live dashboard shows every agent action as it happens. See which agent is working, what it's doing, and how long it's taking. Catch problems the moment they appear — not hours later when someone notices something went wrong.

Live feed of every agent action as it happens

Full history of every past run, searchable and filterable

Problems are flagged immediately with context

One command. Your swarm runs.

no broker setup · no cluster · no build step

docker pull ralphbarendse/swarmwright:latest

read the docs → view on github →