The event log is truth.
State is a materialized view, not a mystery. Runs survive restarts, results re-enter through durable events, and any run can be replayed.
FloofClaw. Tiny code. Sharp claws.
Open source · C11 · runs on your machine
The engine agents run inside.
Not another agent. FloofClaw is the durable control plane underneath agentic work—deterministic where possible, intelligent where necessary, and accountable all the way down.
Quick start
$ git clone https://github.com/FloofLogic/Floofclaw.git
$ cd Floofclaw && make -j4 # build one C11 binary
$ ./bin/fclaw run -h --text "hello" # zero-account mock run560 KiBstripped binary
137 msisolated mock hello
9.1 MiBidle gateway RSS
44.0Kruntime C + headers
Built for the durable part
Do less. Prove more.
Some projects perform intelligent work. Some give an agent tools. FloofClaw owns the part beneath them: what happened, what should happen next, and whether the work keeps succeeding after the demo is over.
Every model call leaves a receipt.
Prompts, responses, metadata, token use, and decisions live as numbered artifacts on disk. The runtime does not ask you to trust a dashboard.
Intelligence is a budget.
Plain code gathers facts and enforces invariants. A model is called when judgment is useful—not because everything has been disguised as a reasoning problem.
One small runtime
Deterministic where possible. Intelligent where necessary.
A floop defines ordered steps. Agents and actions emit events. The deterministic reducer applies those events to state. FloofClaw decides when intelligence is warranted and keeps the durable concerns out of the model's hands.
It does less on purpose: your agents and models keep doing the work; FloofClaw stays accountable for whether that work keeps succeeding.
01
Events arriveUser input, scheduled review, or operation result
event_log.jsonl
02
The floop decidesPlain code first; one inspectable model call when needed
provider_calls/
03
State is reducedActions run, effects are recorded, replay remains possible
state.json
Measured, not vibes
Small is a claim. Here are receipts.
FloofClaw is an ordinary source tree that compiles into one native binary. No container is required for the offline path. These figures were measured on arm64 macOS at v0.31.0 revision a39b487 and are reproducible with make metrics.
Payload and capability
Approximate, rounded project comparison| Project | Payload | vs. FloofClaw | What it can do |
|---|---|---|---|
| SubZeroClaw | ~60 KB | ~0.10x | one skill + one endpoint + unrestricted shell loop |
| FloofClaw | 574 KB measured binary | 1x | durable events, reducers, recovery, scheduling, workers, providers, extensions |
| zclaw | ~850 KB | ~1.5x | ESP32 firmware with Wi-Fi, TLS, GPIO, I2C, schedules |
| NullClaw | ~4.5 MB | ~7.8x | standalone assistant: providers, channels, MCP, voice, tools |
| ZeptoClaw | ~7 MB | ~12.2x | streaming agents, swarms, batch, plugins, containers |
| ZeroClaw | ~33 MB | ~58x | full assistant platform with gateway, dashboard, approvals |
| PicoClaw | ~36 MB | ~63x | Go assistant with MCP, cron, subagents, hooks, Web UI |
| Hermes Agent | ~120 MB installed; ~170 MB with Python | ~209x installed; ~296x with Python | self-improving agent: skills, memory/search, 40+ tools, UIs, gateway |
| OpenClaw | ~87 MB package; ~360 MB installed | ~152x package; ~627x installed | Node platform: integrations, operator UI, plugins, onboarding |
Approximate project payloads supplied for this comparison. Ratios use FloofClaw's 573,768-byte stripped macOS arm64 binary measured at v0.31.0 revision a39b487. Package, installed, firmware, and Python-inclusive sizes are different measures; treat the ratios as order-of-magnitude context, not benchmark parity. Payload is storage or package size, not live RAM.
Tiny code. Sharp claws.
Read the source. Follow the work.
FloofClaw is open source under Apache 2.0. Take a look at the code, run the mock-backed path without an account, and star the repo if the approach is useful to you.