Skip to content

Nucleus: The Open-Source, Self-Improving Software Factory

The open-source, self-improving software factory.

Works with your favourite coding agents:

  • Claude Code
  • Codex
  • Gemini
  • Cursor
  • OpenCode
Coming soon

Our definition of a software factory

A software factory combines deterministic processes, agentic intelligence and human judgment to produce software and improve how it is produced.

Atomic workflows

Issues are split into small workflows

Each workflow finishes one job, then your rules decide whether it moves on.

Workflows Steps Issue Workflow 1 1 2 3 4 Workflow 2 1 2 3 4 Workflow 3 1 2 3 4 Workflow 4 1 2 3 4

Zero trust execution

Each agent gets only the access it needs

Nucleus keeps the credentials, the agent gets a one-time token scoped to the step.

Systems Nucleus Sandbox GitCloudDatabase Holds the keys GitHubCloudDatabase the keys never leave Token one-time Agent credentials nothing to leak, nothing to outlive the task

Runtime evals

Auto-optimised across harness, model and effort

Steps that repeat build a record of time, cost and quality. Each one settles on the cheapest configuration that still hits the quality bar.

Overall 91% 1 Research 97% 2 Draft 95% 3 Review 94% 4 Collate 96% Review 128 runs quality cost bar 90% Model B · medium $0.31 · 94% Model B · high Model A · low

Unified inbox

Every human task in one unified inbox

Decisions, approvals and reviews from every workflow arrive in one place, each with the work behind it attached.

Workflows Inbox You Workflow 1 Decision Workflow 2 Approval Workflow 3 Review Workflow 4 Approval

Governed self-improvement

The factory improves the factory

Evidence becomes a proposal, the proposal is proved against the regression set, and only then is it promoted to govern new work.

RunEvidencePropose PromoteRatifyProve regression set next run only proven changes are promoted

Read next

Docs map

Start here
Overview Factory tour
Improvements
Self-Improvement
Reference
Architecture