Delegatron GitHub

Open source · runs on your own hardware

AI models that
check each other's code.

Describe what you need in plain words. Delegatron plans it, writes it, runs the tests and reviews the result, and goes around again until it passes or the budget runs out.

View source on GitHub How it works

Try the loop · describe a task

  1. PlanWaiting
  2. BuildWaiting
  3. TestWaiting
  4. ReviewWaiting

A simulation of a real run. Press Run and watch the scene follow along.

How it works

Four roles, one loop, clear rules.

Ask one AI model to write code and it will tell you the code works. Delegatron splits the job the way a small team would, so nothing reaches you unchecked.

  1. 01

    Plan

    An architect turns your request into small tasks, each with acceptance criteria and the tests that prove it.

  2. 02

    Build

    An implementer writes the code and its tests as a patch. Nothing touches your project until the patch is checked.

  3. 03

    Test

    The patch runs in an isolated sandbox. Failing tests are the one opinion nobody gets to argue with.

  4. 04

    Review

    A reviewer reads the change and the evidence, then approves it or sends it back with the problems in hand.

The dashboard

See exactly what happened.

Every plan, patch, test report and review is kept, so a run is never a black box. Start runs, follow them live, and resume or retry them later.

Run detail: seven iterations with their test results, review verdicts and cost

Screens from a real run on local models. That run stalled after seven iterations, and the dashboard shows exactly why.

Recorded runs

What it's good at, and what it isn't.

Results from end-to-end runs, start to finish with no human edits. Local costs are electricity and hardware wear.

Built for new code: services, tools, modules and test suites. It isn't a bug fixer for large existing codebases, and a person still has to review what it writes.

Features

Careful by design.

  • Any model, any role

    Local models on one 12 GB graphics card, cloud models, or a mix. Choose per role.

  • Hard limits, always on

    Iterations, spend and time are capped by Delegatron itself, not left to the model.

  • Evidence settles arguments

    Failing tests beat security checks, which beat reviewer objections, which beat opinions. Repeated disputes get one binding ruling.

  • Checks on every patch

    Leaked secrets, new dependencies and imports that won't resolve are caught before a patch is accepted.

  • Proof it stayed local

    Each run records every network destination it touched, so you can show that nothing left the building.

  • Thirty templates

    Web APIs, command-line tools, audits and test suites, with the details filled in when you start.

  • Six languages

    Python, JavaScript and TypeScript, Go, Rust, Java and PHP, each tested in its own sandbox.

  • Terminal, API or dashboard

    The same loop from a command line, a web API with live updates, or the dashboard. Every run gets its own branch.

Open source

Finished, and shared as it stands.

Delegatron is a completed portfolio project by StoriaBrevi. The code is free to read, run and learn from under the AGPL‑3.0.

View source on GitHub
Status
Version 0.3, complete and not maintained
License
GNU AGPL‑3.0, no warranty
Support
None, including installation help
Contributions
Issues and pull requests are not accepted