NullHub is the management console for the NullClaw ecosystem. It installs, configures, supervises, monitors and updates the components around it — NullClaw itself, plus NullBoiler for workflows, NullTickets for task storage and NullWatch for tracing — from a single browser dashboard or an equivalent CLI.
It ships as one Zig binary with a SvelteKit UI compiled into it, so there is no separate web asset directory at runtime. The backend runs in one of two modes: a server with HTTP endpoints and supervisor threads, or a CLI that makes the same calls directly and exits. All state lives under a single directory holding config, instances, binaries, logs and cached manifests.
The engine itself is generic and manifest-driven. Each component publishes a manifest describing how it is installed, configured, launched and health-checked, along with its wizard steps and UI modules; NullHub interprets that rather than hard-coding each component. Requests to component-specific paths are reverse-proxied to the corresponding service, which is how the built-in NullBoiler, NullTickets and NullWatch pages read live data without sending anything to a hosted service.
A Mission Control page runs a deterministic local replay of an agent mission — workflow execution, role-based agents, an induced failure, checkpoint recovery and live telemetry on one screen. It needs no hosted infrastructure or model secrets, and hydrates with real workflow and trace detail when matching local instances are running. The current snapshot, its source fixture and a failed-versus-recovered comparison can be exported as a portable JSON artifact and stored durably for later review.
Features
- Install wizard: manifest-driven guided setup with component-aware flows and automatic linking between dependent components
- Process supervision: start, stop and restart instances, with crash recovery on an exponential backoff
- Health monitoring: periodic HTTP health checks feeding dashboard status cards and live updates over Server-Sent Events
- Config management: structured editors per component, with raw JSON editing as a fallback
- Log viewing: per-instance tailing and live streaming
- One-click updates: download a new version, migrate its config, and roll back on failure, individually or in bulk
- Multi-instance: several instances of the same component run side by side, addressed as component and instance name
- Web UI and CLI: a browser dashboard for operators and a CLI covering the same operations for automation
- Component proxies: workflow, key-value store and tracing APIs are reverse-proxied so their pages render inside the hub
- Flight recorder: run summaries, span timelines, eval results, token usage, cost and error context from local traces
- Mission Control: deterministic local mission replay with checkpoint forking, telemetry and exportable replay artifacts
- OS service: registration as a systemd or launchd service, with status and removal commands
