Direct chat with AI distilled from real human writings. Every answer grounded in actual source texts under enforced retrieval — zero hallucination. Three standalone properties:

• 清晨甘露 · Morning Manna — daily devotional reading + audio. Offline + background playback, lock-screen control, resumable across episodes.
• 细拉 · Selah — converse with 44 spiritual masters spanning 19 centuries (Augustine, John Chrysostom, Brother Lawrence, Spurgeon, 倪柝声, 钟马田 …). Every reply cites the source work + page.
• 岐黄 · Yellow Emperor's Way — converse with 39 TCM masters spanning 4,500 years (黄帝 → 当代国医大师). 498 books, 159 MB corpus, enforced citation.

Specialist conversational agents — language tutoring, citizenship prep, spiritual companion, TCM diagnosis. 7-day free trial, Pro upgrade for continued use.

• 🕊️ Madame Guyon — spiritual companion & inner way guide
• 🎓 English Tutor — 100 real-life scenarios + pronunciation coach
• 🏥 TCM Master — classical diagnosis across 12 master physicians
• 🇺🇸 Citizenship Coach — 128 USCIS questions, bilingual
• 🇨🇳 Chinese Tutor — ABC kids edition · 25 scenarios + quiz
• 🇪🇸 Spanish Tutor — fun bilingual Spanish for kids

The agent runtime that powers everything above, plus the open-source ecosystem and research papers documenting how it all works. Single 23 MB binary, zero cloud dependency, soul-file configured. See Core Innovations + Ecosystem below for the technical surface.

LocalKin is a Go-based AI agent runtime that orchestrates 100+ domain-expert agents through structured multi-round debates, autonomous scheduling, and self-improvement cycles — all from a single 23MB binary.

Thin Soul + Fat Skill — Agent identity in 30-line YAML soul files. Skill logic runs as subprocesses, never sent to the LLM. Token-efficient and injection-resistant.

Zero-Token Heartbeat — Tri-chamber autonomous scheduling (pulse / schedule / idle) coordinates 100+ agents via MQTT without burning tokens. Agents wake on schedule, catch up after restarts, and roll back silent wakeups with zero memory pollution.

Conductor-Driven Swarm Debates — Domain experts debate independently, then update positions after seeing others' arguments. Consensus inertia detection prevents false agreement. Cascade amplification guards catch herding behavior.

Self-Evolving Architecture — SAGE four-step improvement loop (Challenger → Planner → Solver → Critic) with experiential reflective learning. 29+ autonomous improvement cycles, 68+ auto-fixes.

Genesis Protocol — Bare binary + soul file self-bootstraps: hardware probing → skill forging → self-testing → checkpoint resume.

The KinClaw family fits in 4 layers — apps on top, raw macOS bindings at the bottom, kernels in the middle, and a single Mac shell binding it all together. Same SSE event protocol + same .soul.md format across every kernel — one client codebase drives the whole stack, future Linux/Windows shells reuse it without changes.

┌───────────────────────────────────────────────────────────────┐
│  LAYER 4 — apps + chat hub                                     │
│  localkin.ai (Selah · Heal · Manna)   api.localkin.dev (100+) │
└──────────────────────────┬────────────────────────────────────┘
                           │
┌──────────────────────────▼────────────────────────────────────┐
│  LAYER 3 — desktop shell  (Apache-2.0)                         │
│                                                               │
│   ⌘⌥K  ┌──────────────────────────────────┐                  │
│        │  kinclaw-mac v0.4.1  (SwiftUI)   │                  │
│        │  Chat ┊ Cowork ┊ Code            │                  │
│        └────────┬───────────┬─────────────┘                  │
└─────────────────┼───────────┼────────────────────────────────┘
                  │           │
        ┌─────────▼──┐  ┌─────▼──────┐
        │ :5001      │  │ :5002      │
        │ kinclaw    │  │ kincode    │   LAYER 2 — kernels
        │ v1.17.0    │  │ v0.10.0    │   (Apache-2.0 / MIT)
        │ 5 claws    │  │ 10 tools   │
        └────────┬───┘  └────────────┘
                 │
┌────────────────▼──────────────────────────────────────────────┐
│  LAYER 1 — Pure-Go macOS bindings  (KinKit, MIT)               │
│  sckit-go · kinax-go · input-go · kinrec                      │
└───────────────────────────────────────────────────────────────┘

Repos:

Project	Description
ollamadiffuser	Local AI image generation, zero cloud dependency
localkin-service-audio	High-performance local STT & TTS services
kinclaw	macOS computer-use agent — 5 claws + floating chat UI + voice. Agent operates your real Mac. 67.3% on macbench v0.1 (first reference run, 2026-05-08).	Apache-2.0 · v1.17.0
kinclaw-mac	Native macOS Spotlight shell for the kernel family. ⌘⌥K → Chat (100+ cloud agents) · Cowork (kinclaw 5 claws) · Code (kincode + repo).	Apache-2.0 · v0.4.1
kincode	AI coding assistant, 10MB single binary. HTTP+SSE server mode for desktop shells.	MIT · v0.10.0
macbench	First publicly published macOS-native computer-use benchmark. 369 task slots, 15 categories, agent-agnostic Go runner. Inspired by OSWorld; adapted for the macOS app surface OSWorld can't reach.	MIT · v0.2.0
notebooklm-go	Unofficial pure-Go client + CLI for Google NotebookLM's internal batchexecute RPC. Reverse-engineered. List notebooks, add PDF/URL/text sources, generate Audio Overviews / Mind Maps / Reports / Deep Research. Powers LocalKin's content publishing pipeline.	Apache-2.0 · v0.2.2

The libraries that power 4 of KinClaw's 5 claws. Each is independently usable in any Go project that needs to drive macOS at the framework level — no Xcode, no cgo, no Swift bridge.

See the Embedded Dylib paper for the distribution pattern these all share.

The other 4 claws are macOS-bound. The web claw is cross-platform and arguably the most-used in real flows, because most modern productivity lives in browser tabs — Gmail, Linear, Notion, GitHub, Booking, Airbnb, Google Flights, the LocalKin family's own apps (localkin.ai, faith.localkin.ai, heal.localkin.ai, api.localkin.dev).

KinClaw's web claw is 3 tiers, picked by Pilot based on what the task actually needs — cheap-and-fast first, expensive-and-flexible only when nothing else works:

The most important operational doctrine in pilot.soul.md — go straight to result URLs, skip GUI clicking entirely. Covers ~80% of real flows.

Why this beats GUI puppeteering for an LLM agent: calendar pickers (clicking "previous month" 30 times to land on July is doomed; ?checkin=2025-07-08 lands instantly), faceted filters (URL params take 4 dimensions atomically), cookie banners + modals (skipped entirely), SPA accessibility gaps (React tree without AX labels — URL bypasses the DOM).

When URL-first can't reach (page needs JS to render, content lives behind a click, you need a screenshot of the rendered DOM), Pilot drops to the web skill — a Playwright-driven single-shot.

web url=X
web url=X click=".btn" type_text="hello"
web url=X js="document.body.innerText"
web url=X screenshot=true

One call = one Playwright session, ~3s cold start, returns immediately. For 99% of one-off web tasks: scrape a page, run a snippet of JS, fill+submit a form, grab a screenshot.

For genuinely complex flows — login + navigate + extract, fill + submit + verify, search + sort + dig three pages deep — KinClaw wraps browser-use (the 91K★ OSS that is the de-facto LLM-driven browser agent library) as a single-argument super-skill:

browser_session task="登录 GitHub 找我未读 PR"
browser_session task="去 weather.com 输 zip code 95014 拿一周预报表"
browser_session task="open my Linear, find tickets assigned to me this week, summarize"

What you get under the hood:

• Persistent session — login state, cookies, localStorage all survive across steps
• DOM element numbering — browser-use auto-numbers every interactive element so the LLM can refer to "click element 17" instead of fragile CSS selectors
• Visual reasoning — screenshot + numbered elements fed back to the planning LLM each step
• Cross-page flows — natural to chain login → navigate → extract; the inner LLM plans steps, you just give it the goal

Trigger rule (from pilot.soul.md): the task description has 2+ interaction verbs (login + navigate + extract / fill + submit + verify / search + sort + dig).

Cost-aware: browser_session burns LLM tokens per planning step (~$0.05–0.15 Claude per 5-step task). Pilot's soul says explicitly: don't reach for browser_session when web would do. The tier system isn't just capability — it's economics.

Most LocalKin's own products are web — Selah, Heal, Morning Manna, the 100+-agent chat hub. When Pilot drives a LocalKin user flow, the web claw is doing the lift. All three tiers are cross-platform (shell + Node.js, no macOS framework dependencies) — future kinclaw-pal (Linux/Windows shell) inherits the whole web stack with zero rewrite. Only the 4 macOS-bound claws need platform-specific rebinding. The web tier is what makes the Linux/Win story viable.

First reference run (2026-05-08):
  kinclaw v1.15.0 + Kimi-K2.5(cloud) on macbench v0.1
    IMPLEMENTED:  101 / 150  =  67.3%
    STRICT:       101 / 369  =  27.4%   (stubs count as fail)

For context, Anthropic Computer Use scores ~38% on OSWorld (Linux desktop). macbench measures a different surface (macOS native), so the numbers aren't directly comparable, but the methodology + scoring discipline are the same.

xlang-ai/OSWorld (NeurIPS 2024) became the de-facto standard for desktop computer-use agents — but it benchmarks inside an Ubuntu/Windows VM. Nobody had published a comparable benchmark for macOS native apps. macbench fills that gap with 369 task slots across 15 macOS app categories (Finder · Safari · Mail · Notes · Calendar · Reminders · Settings · Terminal · Pages · Numbers · Keynote · Music · Photos · Maps · multi-app), an agent-agnostic Go runner (any binary that takes a prompt + drives macOS plugs in via -agent + -agent-args template), and per-task PID-snapshot isolation that preserves any pre-existing user app state.

v0.1 ships 150 implemented + 219 stubs (real prompts, no setup/eval scripts yet); fill rate over v0.2 → v1.0 is roughly 30-50 stubs/month. License MIT.

100+ specialized agents across expert traditions, each grounded in real source texts:

• Spiritual · 44 masters spanning 19 centuries → faith.localkin.ai (Augustine, John Chrysostom, Brother Lawrence, Spurgeon, 倪柝声, 钟马田 …)
• Traditional Chinese Medicine · 39 masters spanning 4,500 years → heal.localkin.ai (黄帝 → 当代国医大师, 498 books / 159 MB corpus)
• More domains in active development — engineering, quant finance, education, language tutoring, design, game dev, spatial computing, and growing.

All 11 papers are on Zenodo with permanent concept DOIs (CC-BY-4.0). Each PDF bundles EN + 中文 in a single document, generated directly from the canonical Markdown source. Concept DOIs auto-resolve to the latest version, so citations stay correct across revisions.

📧 contact@localkin.ai