Architecture

The ecosystem is a set of small services with a strict division of labor, joined by plain JSON over HTTP. Understanding two things — the layer split and the manifest contract — explains most of how it fits together.

The layer split#

The core stack separates three concerns that most agent platforms fuse into one process:

            NullHub  (control plane: install, supervise, update, dashboard)
               │
   ┌───────────┼───────────────┬──────────────┐
   ▼           ▼               ▼              ▼
NullTickets  NullBoiler     NullClaw      NullWatch
 (truth)      (policy)     (execution)   (telemetry)
   ▲             │             │              ▲
   └── claims ───┘             └── spans ─────┘

• NullTickets owns truth. What work exists, its state machine, its history. It stores the graph and the events; it never runs agents or decides scheduling.
• NullBoiler owns policy. What runs, when, on which worker — routing, concurrency, retries, fan-out, checkpoints. It polls the tracker, claims tasks, and dispatches.
• NullClaw owns execution. Providers, tools, sandboxing, channels. It does the work and reports back.

Each layer is optional. One agent loop plus NullTickets is a valid stack; so is NullClaw alone. Around the core: NullWatch records what happened (spans, evals, cost), NullPantry holds what the team knows (permission-aware knowledge and memory), and NullDesk puts a human at the pass — reviewing every plate before it leaves the kitchen.

The manifest contract#

NullHub is not hard-coded to its components. Each managed component publishes a nullhub-manifest.json describing its installation, configuration schema, launch command, health checks, wizard steps and UI modules. The hub is a generic engine that interprets manifests — components export theirs with --export-manifest, and bootstrap config from wizard answers with --from-json (NullWatch, NullTickets and NullBoiler all ship both).

This is why the managed set is exactly NullClaw, NullBoiler, NullTickets and NullWatch: those are the four that publish manifests today.

Protocols#

No message broker is required anywhere in the core; MQTT and Redis Streams are optional NullBoiler dispatch paths.

Design rules#

The same decisions recur across every repo:

• One static binary per component. No runtime dependency on Node, Python, or provider SDKs. NullHub even embeds its Svelte dashboard into the binary via @embedFile.
• Vendored persistence. Where a database is needed, it is SQLite compiled in — no external DB to run. NullWatch skips even that: append-only JSONL files.
• Local-first state. Everything under a home directory: ~/.nullhub, ~/.nullclaw, ~/.nulltickets, ~/.nullboiler, ~/.nullwatch. No hosted service exists to sync to.
• Interfaces over plugins. NullClaw makes everything a swappable vtable interface — providers, channels, tools, memory engines, tunnels, runtimes, peripherals. NullPantry moves the same idea to compile time: engine profiles (nullclaw/minimal/full/custom) build in only the backends you chose.
• Loopback by default, tokens for the rest. See Security.
• Zig 0.16.0, CalVer, shared CI. Every repo pins the same toolchain and releases through NullBuilder's reusable workflows.

Next: the seams you can build against — Integrations.