A run should survive a crash.
Work is recorded as it happens, so interruption does not silently turn into repetition or lost context.
← Selected work
Open-source agent runtime
Nexus
A small TypeScript runtime for AI work that needs to pause, survive, respect limits and explain itself later.
The premise
When software can act, somebody still owns the consequence.
Most agent demonstrations concentrate on whether a model can complete the task. Operational software has to answer a different set of questions.
What happens when the process stops halfway through? Who approves the risky action? Where does the cost stop? Can another person understand what happened tomorrow?
Nexus began from the belief that these questions belong inside the run, not scattered across disconnected tools and good intentions afterwards.
- Role
- Creator & maintainer
- Language
- TypeScript
- Licence
- MIT
- Release
- Public npm package
Three decisions
The design is easier to understand without the jargon.
Risky actions should pause, not pray.
Human approval is part of the state of the system. A person can approve, edit, answer or refuse before the tool acts.
Every result should explain itself later.
The run leaves a durable record that can be inspected, replayed and compared rather than reconstructed from memory.
Public evidence
Small enough to inspect.
Serious enough to test.
171automated tests
0runtime dependencies
v0.2public release
The repository includes the runtime, an open run format, examples, tests and operating guidance. The boundary is explicit: Nexus provides the governance core; production teams still own identity, network and infrastructure controls.
Inspect the source ↗The interesting part is always the problem underneath.
If you have one that crosses commercial and technical boundaries, start there.
Bring the problem ↗