FloofClaw
tiny code. sharp claws.
Not another agent. FloofClaw is the engine agents run inside — a durable control plane that decides when intelligence is needed, and handles everything else without pretending it needs an LLM.
floofclaw.com
private beta
447 KBstripped binary
7 msfull durable run
9.7 MBRSS, live prod bot
27Klines of C11
Small is a claim. Here are receipts.
| engine | written in | ships as | memory | speed |
|---|---|---|---|---|
| floofclaw | C11 | 447 KB binary | 9.7 MB live¹ | 7 ms full run |
| nullclaw | Zig | 678 KB binary | ~1 MB idle | <2 ms start |
| picoclaw | Go | MB-class binary | 10–20 MB | <1 s boot |
| nanoclaw | TypeScript | Node + Docker | unpublished | unpublished |
| openclaw | TypeScript | Node ≥22 + npm tree | unpublished | >500 s boot² |
¹ resident memory of a real production bot with Discord + IRC connected, not an idle hello-world. ² competitors' numbers from their own READMEs (0.8 GHz reference hardware); ours measured 2026-07-17, arm64 macOS.
Deterministic where possible. Intelligent where necessary.
Some projects perform intelligent work. Some give an agent tools. FloofClaw is the runtime underneath: it holds the durable concerns, wakes on a schedule, gathers the facts with plain code, and spends exactly one model call deciding whether anything deserves your attention — whether the workload was ten records or ten thousand.
The event log is the source of truth. Every LLM call is its own inspectable run on disk — the prompt, the response, the token bill, the decision. Replay any of it. It does less on purpose: your agents and models keep doing the work; FloofClaw is accountable for whether the work keeps succeeding.
Floof Logic LLC · [email protected]