A build space where progress can listen.
Agents can write a great deal of software very quickly. What they cannot do on their own is tell you whether any of it was worth writing. Mirasim puts both halves in one space: the agents that build, and the simulated users who try the result and say where it fails.
Download for macOSAgents build.
Up to six agents work in parallel panes, each in its own git worktree — Claude Code, Codex, and Mirasim's own. A full editor, terminal, file tree and artifact store sit in the same window, so the work never leaves the space it started in.
Simulated users try it.
Hand a build to a simulated person and they work through the job on a real browser, phone or shell — clicking what they'd click, giving up where a real user would give up. You pick who tries it the way you'd pick a test group.
Evidence decides.
Every run comes back scored against a fixed rubric, and every number links to the moment that produced it. Keep the change or roll it back on what the runs say — then go round again.
Why a loop, and not a faster editor
The bottleneck in building a product stopped being the typing some time ago. It is knowing which of ten plausible directions a real person would actually follow — and that answer normally arrives weeks later, from users you have already shipped to.
A simulated user closes that gap to minutes. Not because a simulation is the same as a launch, but because most of what you need to learn early is coarse: the step nobody finishes, the wording nobody understands, the button nobody finds. Those show up in a simulated run, they show up cheaply, and they show up before the work is sunk.
So the unit of progress here is not a commit. It is a round: build, put it in front of people, read what happened, keep what won.
Everything shown is real
Every recording and screenshot on this site comes from the shipping product on a real workspace. Nothing here is a mock-up or a re-enactment, even when a real capture is the less flattering one.
Evidence over opinion
An evaluation is only worth something if it can be wrong. Runs are scored on a rubric fixed before the run, and reports link back to the traces they came from.
Your stack stays yours
Your subscriptions, your API keys, your models — swap them mid-build, add MCP servers, run models locally. Nothing forces you onto one provider.
We fund the build.
$100 in build and evaluation credits for every early builder, and $3,000 more by application, so every token goes into the product rather than the bill. Credits only — we take no stake in what you build. Applications are open.
See the Build Fund →Questions about the product, the Fund, or working together: mirofish@shanda.com. We read everything that arrives there.