SwarmCoder
Overview How it works Walkthrough Agents Principles Status About
APACHE 2.0 GITHUB ↗ ● OPEN SOURCE
RELEASED · APACHE 2.0 · UNDER ACTIVE DEVELOPMENT

Where this is today

The code is published under Apache 2.0 at github.com/swarmcoder ↗. It is substantially built and still under active development, and this page says plainly what is and is not proven yet.

WHAT EXISTS
✓ The requirement and story model, with executable checks
✓ Document intake and requirement extraction
✓ The fixed workflows and their machine-checked invariants
✓ The swarm scheduler and worker sandboxing
✓ The verification harness and candidate archival
✓ Full session tracing and persistence
✓ The operator console described on these pages
WHAT IS HONESTLY NOT PROVEN YET

The full pipeline has so far been driven by tests and simulated model servers far more than by live ones. The claim that a swarm of small local models beats a single frontier pass on real work in a real repository is the thesis this project exists to test.

It is well supported by the published research but not yet by our own numbers. When we have those numbers they will be on this page, including the ones that disappoint.

It runs on two machines

TOKEN SERVER
A DGX Spark class box serving tokens and nothing else. Any OpenAI compatible endpoint works for the local roles, and the worker model is a configuration choice that is deliberately interchangeable.
ORCHESTRATOR
An ordinary workstation running the orchestrator, the sandboxes and the builds, on your network and against your repository.

What SwarmCoder is not

Not an IDE plugin
It is a standalone system with its own console. Chat is available everywhere as a companion but is required nowhere.
Not configurable into a different process
There are no plugins, no agent graphs and no workflow language, and this is deliberate.
Not fast
It is built for long unsupervised runs, so start it in the evening and judge the deliveries in the morning.
Not a hosted service
It runs on your hardware, on your network and against your repository.
Not a way to avoid deciding
It asks you exactly twice per story and it expects real answers from you.
What it is
One process, followed every time, with a full record of what was done. See the cycle →
APACHE 2.0
Open source, under the same license as ZeroZ4j, the framework it is built on.
GITHUB ↗
The code, issues and releases live at github.com/swarmcoder.

If you want an agent that will do anything you ask in any order, there are excellent ones out there already and you should probably keep using them. SwarmCoder is for the case where you would rather have one process that is followed every time and leaves a record you can audit.

The code is at github.com/swarmcoder ↗. Questions: franzschoning.com ↗
SwarmCoder

Agentic software development with full traceability, built on ZeroZ4j by Franz Schöning, Principal Enterprise Architect.

swarmcoder.dev GitHub ↗ zeroz4j.com ↗ franzschoning.com ↗ ● Open source · Apache 2.0
© 2026 Franz Schöning. Released under the Apache License 2.0.