KinClaw
The self-fissioning lobster. Breeds its own swarm on demand.
可以裂变的龙虾 — 根据需求自己造龙虾群。
KinClaw is a computer-use agent — primary target macOS, also runs on
Linux and Windows as of 2026-05-12. It sees your screen, understands
your UI semantically, clicks, types, and — the part no one else has —
reproduces on demand via three primitives:
- Soul Clone (
pkg/clone) — duplicate a specialist into N
parallel workers with small per-clone divergence.
- Skill Forge (
pkg/skill forge) — when an existing skill
can't handle a task, KinClaw drafts, writes, registers, and tests
a new one, then retries.
- Sub-agent dispatch (
spawn skill) — fork-exec a specialist
child on a different brain (researcher / eye / critic / coder ship
in-box; hierarchical, kernel-capped at depth 1).
Single binary, ~18 MB, Go 1.25+, Apache-2.0 licensed. Project page:
localkin.app/kinclaw.
🙏 Help wanted — Linux + Windows testers
The agent wrote 2,042 lines of Linux + Windows port code from API
docs without hardware to test against on 2026-05-12. Builds are
green; runtime behaviour is unverified. If you have a Linux box or
a Windows machine and 10 minutes to spare, grab a prebuilt binary
from v1.17.0
(v1.18.0 shipped without binaries) and try a few smoke commands. See TESTING.md for
a 3-command quickstart + report template. Track Linux progress in
#1, Windows in
#2. Both ✅ and
❌ reports are useful.
Primary target: macOS (tested daily-driver).
Linux Phase 2-5 + Windows Phase 6 both landed 2026-05-12 — code
complete, awaiting community runtime testing
(#1 Linux,
#2 Windows):
| Layer | macOS | Linux | Windows | Backend (non-mac) |
|---|
screen claw | ScreenCaptureKit | ✅ | ✅ | Linux: grim / scrot / xrandr · Windows: PowerShell + System.Drawing |
input claw | CGEvent | ✅ | ✅ | Linux: xdotool / ydotool · Windows: user32.dll P/Invoke + SendKeys |
ui claw — window-level | AX | ✅ | ✅ | Linux: xdotool / wmctrl · Windows: user32.dll GetForegroundWindow |
ui claw — a11y tree | AX | ✅ | ✅ | Linux: AT-SPI 2 via godbus · Windows: UI Automation (UIA) |
record claw | ScreenCaptureKit | ✅ | ✅ | Linux: ffmpeg (x11grab / pipewire) · Windows: ffmpeg (gdigrab) |
location skill | corelocationcli | ✅ | ✅ | Linux: gdbus + geoclue2 · Windows: Windows.Devices.Geolocation (ipapi.co fallback) |
cerebellum/*.sh library | 16 cats (478 actions) | 4 cats (linux-files/apps/settings/clipboard) | 4 cats (windows-files/apps/settings/clipboard) | shell + POSIX tools / PowerShell |
pilot soul | 24 skills | 23 skills (no app_open_clean) | 23 skills (no app_open_clean) | — |
Daily-driver souls:
Same starter lobster for everyone. Every user's swarm is unique
after a month.
KinClaw grew out of the earlier localkin runtime (a minimal
embodied-AI microkernel, ~2,300 lines). This repo is that same
skeleton with the five claws bolted on: screen, input, ui,
record, and web (Playwright). On macOS the first four use
ScreenCaptureKit / CGEvent / Accessibility API / kinrec via their own
zero-cgo KinKit libraries; the Linux + Windows ports swap those for
native backends (see feature-parity table above) without changing the
skill API.
🆕 The grep-routed agent stack (2026-05-11, paper #11)
Since 2026-05-11 KinClaw ships a four-layer NL → action router
(skills/kinthink/) on top of a multi-platform skill library
(skills/cerebellum/ — 16 macOS cats / 478 actions, 4 Linux cats,
4 Windows cats; each backend swaps in transparently per runtime.GOOS).
For prompts that match a known canonical operation — file rename,
note create, calendar event, settings toggle, web fetch, … — the
router skips the LLM entirely:
$ kinclaw -soul souls/macbench.soul.md -exec "rename foo.txt to bar.txt"
★ matched : 001-finder-rename (tf-idf=3.43, 20ms)
★ substituted: cerebellum 'finder rename ~/Desktop/kinbench/foo.txt ~/Desktop/kinbench/bar.txt'
★ router : 65 ms
ok: rename /Users/jackysun/Desktop/kinbench/foo.txt -> /Users/jackysun/Desktop/kinbench/bar.txt
★ TOTAL : 53 ms (router 20 ms + exec 33 ms)
real 0.56 # ← 0 LLM tokens, 50-500× faster than agent loop
On the macbench v0.2 benchmark (379 tasks):
| Configuration | Pass | Time | Tokens |
|---|
| LLM-only baseline (paper #10) | 30.4% | 107 min | Full |
| Reference verifier (no LLM) | 84.3% | 22 min | 0 |
| kinthink + cerebellum (this stack) | 48.0% | 76 min | 0 on Layer-0 hits |
Opt in by setting cerebellum.grep_route: true in your soul. See
paper #11 for
the architecture + measurement.
Quick start
# The souls live in the repo, so clone it — `go install …@latest`
# alone gets the binary but not souls/pilot.soul.md.
git clone https://github.com/LocalKinAI/kinclaw && cd kinclaw
go install ./cmd/kinclaw
# Default pilot runs Kimi K2.6 via Ollama Cloud. Sign in once:
ollama signin
# The pilot soul — the generalist that drives your Mac
kinclaw -soul souls/pilot.soul.md
# Then ask it something like:
# > "What app is in front? Click the Save button if there is one."
Want a specialist instead of the generalist pilot? KinClaw ships four
focused souls; pilot dispatches to them via spawn, but you can also
launch them directly:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
