A software build session turns one message into a whole build. Instead of one model answering turn by turn, Elzi plans the work, splits it into tasks, hands them to several nodes at once and writes the files into your project folder. The app calls these swarm builds. You start one by choosing "Elzi · Build" as the chat's model and describing what you want built.
Starting a build
- Open a chat in the project folder the build should write to.
- In the model picker choose "Elzi · Build" (plan and fan out a build across the pool).
- Describe what you want built. The build runs on your swarm pool, or on a strategy you marked "Use for swarm builds".
What happens during a build
- Open: the brief is sliced into parts.
- Ask: the swarm asks you once, up front, about decisions only you can make.
- Research: one read-only lane per part, in parallel across your nodes.
- Plan: tasks, the files each task owns and the order they depend on.
- Build: large tasks are split into smaller ones, and the nodes write the code.
- Integrate and repair: the app is started, probed, and what the checks find gets fixed.
A run panel shows each phase live. Every phase is also written to an event log in the project.
Roles and nodes
A strategy assigns nodes to roles: Planning (reading the request, asking questions, researching and writing the plan), Build (writing the code for each task), Backend, Frontend, Testing (running and checking the result, and fixing what the check finds) and Chat. Each node has a weight; a higher weight gets a bigger share of the tasks. LM Studio models loaded on your Macs can join a build too.
Limits
- OpenAI-compatible endpoints you add yourself are chat only for now; swarm builds can't use them yet.
- Each Mac's MLX engine serves one model at a time, so a build reaches a Mac's local model only through that Mac's single engine.
How it differs from agent sessions and benchmarks
| Agent session | Software build | Benchmark run | |
|---|---|---|---|
| You start it with | A chat in a project folder | A chat on "Elzi · Build" | The Benchmark view |
| Who drives | You, turn by turn (or a schedule, for Agent Work) | The swarm, from one brief | A frozen task, the same for every model |
| Models | One model or node answers | Many nodes share the work | The model or nodes you want to measure |
| What you get | The conversation and the files it changed | A planned, multi-agent build written into the project | A score with its evidence, which you can compare |
Facts on this page were checked against our own test records on 6 November 2026.