Every agent at once
One task goes to every coding agent you have available, CLI agents and API endpoints alike, in parallel.
Runs on your machine, with the agents you already have
One task. Every agent. One verified winner.
Cyberine sends one task to every coding agent you have, runs each in its own git worktree, and keeps the diff that actually applies and checks out.
A real run, recorded 2026-09-25
- kept
- loop-deepseek-build verified
- diff
- 1 file(s), +4/-1
$ cyberine run --cheap "Fix the off-by-one bug in src/chunk.ts"
dispatch 4 implementers, each in its own git worktree
loop-deepseek-build done
loop-commandcode-free stuck, cancelled by the stuck guard
loop-mimo-build done
loop-openrouter-build done
judge DiffJudge applies every diff in a throwaway worktree
loop-mimo-build diff applies cleanly (2/2)
loop-openrouter-build diff applies cleanly (2/2)
loop-deepseek-build diff applies cleanly (2/2)
diff 1 file(s), +4/-1 from loop-deepseek-build
apply cyberine apply 81cbc07c
Evidence over confidence
Stop trusting one model’s word
An agent saying “tests pass” is a claim. Cyberine checks the claim before anything touches your tree.
What a worker says
Tests pass.
What the judge does instead
- apply the diff in a throwaway worktree
- git apply succeeds
- tsc --noEmit passes
both gates passed (score 2/2)
Isolated worktrees
Each worker that writes code gets its own detached git worktree off HEAD. They cannot collide, see each other, or touch your live tree.
Hard spend cap
CYBERINE_DAILY_USD_CAP is checked before any worker starts. Subscription seats, a logged-in Claude or Codex plan, add no marginal cost per run.
Judged on evidence
The judge applies every diff and runs the project’s typecheck. The one that survives wins, not the one that sounds sure.
Four stages, the same every time
From one prompt to a landed diff
The recorded run, as a git graph. Scroll drives it.04 Land
- 01
Dispatch
Your task fans out to every available agent. Loop workers and CLI workers start from the same commit at the same moment.
- 02
Build
Each one edits code in its own worktree. None of them can see another, and a worker that stalls is cancelled by the stuck guard without stopping the rest.
- 03
Judge
Judging starts only when the last worker has stopped. Every diff is applied and checked: git apply, then the project’s typecheck. Two gates, two marks per candidate.
- 04
Land
You apply the winner with one command. The surviving diff is recorded in the run’s run.json first, so nothing reaches your tree on a worker’s say-so.
Picking the winner
Four judges, one of them a compiler
For code, DiffJudge is the default: the verdict comes from git and the compiler, not from another model.
DiffJudge default for code
Applies each candidate diff in a throwaway worktree and gates on git apply plus the project's typecheck.
git applythentsc --noEmit
AdversarialJudge
Sends several skeptics to refute each answer.
LlmJudge
A single model reads the results and compares them.
ConsensusJudge
Picks the answer most agents agree on.
Hover or focus a judge to watch how it picks. All four judges, in detail.
Surfaces
One engine, four ways in
All four share one run store and the same judges.
Command line
One command, cyberine run "<task>", with JSON out if you want it.
Fleet TUI
Every run across every project, in the terminal.
Web dashboard
A local dashboard, served on 127.0.0.1.
MCP server
Call the fleet from another agent.
one run store, the same judges
Workers
The agents you already have
Cyberine dispatches to whatever is present on your machine.
CLI workers
A real headless coding agent, spawned as a subprocess, inside a host sandbox when one is available.
- Claude Code
- Codex
- Muse
- Command Code
- Cursor Agent
- Aider
Loop workers
A tool-calling loop against any OpenAI-compatible endpoint.
- DeepSeek
- OpenRouter
- Xiaomi MiMo
- Ollama
- LiteLLM gateway
- any OpenAI-compatible endpoint
Run your first race.
Install it, point it at a repo, give it a task.
run 81cbc07cangle: the racerecorded 2026-09-254 workers3 diffs applied1 kept$0.10 spent