SwarmCoder is a multi-agent coding system that follows one fixed software development lifecycle every time it runs. It uses expensive frontier models only where a task genuinely needs that level of judgment, does the bulk of the work with swarms of cheap local models on your own hardware, and writes down every decision it makes along the way so that you can go back and check it later.
You make two decisions per story and the system does the rest.
Everything happens on one board. A story appears when the planner suggests it, you decide whether it is real work, the swarm builds it, and then you judge whether the delivery is what you asked for. Nothing disappears off the board and nothing gets parked somewhere you have to remember to go and look.
For any line of code you can see which worker wrote it, which candidate it was part of, which test proved it and which requirement asked for it, right back to the sentence in the document it came from. And you can still query this months later, in either direction.
The token heavy part of the work, i.e. the loops that read through your whole repository, runs on your own hardware. The cloud based roles only ever see typed summaries and diffs, never a raw worker session, and whatever cloud spend remains is capped per run.
Every requirement is bound to an executable check, so it is only marked as implemented when there is evidence (a commit and a passing test) and not because somebody said so. If you edit the wording of a requirement, the evidence is marked as stale and you can see that on the board.
On hardware you own, running eight workers on the same task costs roughly the same as running one, because the model weights are read out of memory once per step regardless of how many sequences are in flight. So if seven of the eight attempts get thrown away, that is fine, that is basically how the method is supposed to work. You can only afford to think this way when the marginal cost of a token is your electricity bill.
SwarmCoder is a multi-agent system, and which model an agent runs on is essentially a cost decision made per task. Every agent runs on the cheapest model that can do its particular job well, whether that is a frontier model, a mid-tier one or a local one. The rule of thumb is simple enough: you need a frontier model where there is no test sitting in front of the output to catch a mistake.
Workers get eight attempts and a verification gate, so they can run as swarms of local models. The judgment roles, on the other hand, get one shot at an artifact that nothing downstream can check mechanically, and that is where the frontier spend goes. Keep in mind that this spend is small, structured and capped per run.
Meet all fifteen agents →The price you pay for frontier tokens today is a market share price and not a cost price, and most agentic coding platforms are built to consume more and more of them with every release. SwarmCoder is designed the other way around, and it was designed that way before the repricing rather than after it.
Most tools that run agents in parallel simply hand you a row of tabs and twenty diffs to read through yourself. SwarmCoder picks the winner mechanically: the tests are written first in a place the workers cannot touch, the cheap filters (does it parse, does it compile) run before the expensive ones, the survivors are grouped by how they actually behave, and only then does a judge rank a handful of genuinely different diffs against the real test results.
Exactly one candidate wins, and alternatives are never merged together. The losing candidates are archived along with their evidence rather than deleted, so you can always go back and see what else was tried.
See how selection works →SwarmCoder is for organizations that need proper traceability on their software development, i.e. one process that is followed every time and leaves a record behind, rather than an agent that reinvents its way of working on every run. It runs on your own hardware, on your own network and against your own repository. If you are a solo developer who is happy with a chat based coding assistant, this is probably not for you, and you should not change what already works.
Agentic software development with full traceability, built on ZeroZ4j by Franz Schöning, Principal Enterprise Architect.