nexusopen source · MIT nexusopen source · MITEvery run: a persistent, auditable, replayable file, with OpenTelemetry traces of every step. Zero dependencies. No cluster, no cloud, no vendor.
Try the live demo¹ measured July 2026, methodology
Every step persists before execution continues. nexus resume picks up exactly where a crashed run stopped, hours later, any process.
Mark a tool requiresApproval, the run pauses durably. Policy rules auto-decide the easy cases. Everything journaled, with who decided.
Token and USD caps checked before every model call. The $81k-in-a-week story (real, 2025-26) becomes a $50 halt and a Slack ping.
Finished runs re-execute offline from their own record, or diff against a new model. Incidents become regression tests.
One call ships OpenTelemetry spans (invoke_agent, chat, execute_tool) with token usage to Datadog, Grafana, Langfuse, any OTLP backend. Zero deps, no vendor SDK.
Tag a run with a customerId and it rides the record, journal, webhooks, and every span. "Whose runs doubled the bill?" becomes a filter, not a forensics project.
The Open Agent Run format (public domain) makes every run auditable, resumable, and replayable, by any tool, forever. That's the difference between a framework feature and a standard.
How it works → Enterprise playbooks & case studies → Star it on GitHub