Сильнее ChatGPT? OpenHuman и local Skills набирают обороты

Весна 2026: в dev Twitter и tech feeds одно имя — OpenHuman. С 0 до 17k+ GitHub stars меньше чем за три месяца; сотни комментариев на Product Hunt и Hacker News. Первая реакция: «ещё один ChatGPT wrapper?» После кода и docs: не совсем — differentiator в «local Skills + Memory Tree + readable wiki», не model itself.

Фокус статьи — Skill system: extension agent, сравнение с ChatGPT GPT, beta + runtime migration, почему Apple Silicon developers deploy stack на Mac mini M4 Cloud Mac.

Почему OpenHuman buzz в Silicon Valley

ChatGPT отлично ведёт диалог; три chronic pains у developers:

• Cold start : new chat = zero project/mail/Slack context ;
• Vendor lock-in : Custom GPT tied to OpenAI cloud account ;
• No audit : нельзя review «memory» как code.

OpenHuman (GPL-3.0, Rust + Tauri):

• Memory Tree : Auto-fetch Gmail/Slack/Notion → local wiki (Memory OS) ;
• Local Skills : install/uninstall, plaintext on machine ;
• UI-first : minutes onboarding, не terminal YAML hell.

Narrative: «model can change; workflow knowledge and extensions stay local.» Дополняет personal context vs ChatGPT Memory — Memory remembers ; Skill acts.

Что такое Skill OpenHuman

Не «ещё один ChatGPT Plugin» — extension pack в agent workspace:

• manifest metadata : name, description, tags, version ;
• SKILL.md body : plaintext workflow, inject at inference.

Official registry: tinyhumansai/openhuman-skills. Discover → install → uninstall ; open SKILL.md in editor to audit boundaries.

Injection в «brain» agent

По Skill injection PR (May 2026):

User message
    ↓
Skill matcher (@ explicit + keyword/tag heuristics)
    ↓
read_body() → SKILL.md matched Skill
    ↓
render_injection (8 KiB cap, truncate if over)
    ↓
Agent::turn — LLM + Skill instructions + Memory Tree
    ↓
Built-in toolchain (filesystem / git / web / integration APIs …)

## Available Skills lists Skills ; real work — matched SKILL.md body, versionable SOP for model.

Текущее состояние: Skill runtime migration

May 2026 = transition period.

• Early: QuickJS sandbox for JS in Skill packages ;
• GitBook + community: QuickJS/rquickjs layer removed ;
• Now : metadata catalog + SKILL.md prompt injection, not full third-party executable plugin runtime ;
• Runtime rebuilding — treat as beta.

Install Skill today = mostly structured instructions + workflow templates with built-in toolchain — not arbitrary closed binary. Security win for sensitive teams ; adjust expectations if you want «one-click any API plugin».

OpenHuman Skill vs ChatGPT Custom GPT

«Сильнее ChatGPT» для composable, auditable, offline-capable agent workflows — да architecturally. Pure conversational IQ или cutting-edge multimodal — cloud frontier models often lead. OpenHuman sells agent infrastructure, not bigger-parameter LLM.

Skill + Memory Tree: 1 + 1 > 2

• Skill: «how to handle customer escalation ticket» ;
• Memory Tree: «customer Slack/mail summary last 30 days» ;
• Model Routing: hard reasoning → strong model, embedding → local Ollama ;
• Coder toolset executes engineering steps from Skill instructions.

Custom GPT can have Instructions, but auto-ingest Slack threads into local wiki — structural OpenHuman advantage for personal/small-team agents.

openhuman-skills: community extension entry

• src/core/ — shipped Skills (TypeScript + manifest) ;
• docs/SKILL_SPEC.md — writing spec ;
• Pipeline: TS compile → esbuild bundle → registry ;
• Default catalog = this GitHub repo (VITE_SKILLS_GITHUB_REPO overridable).

Fork, SKILL.md per SPEC, self-hosted catalog — not wait for official store. Differs from OpenClaw (gateway orchestration vs desktop personal agent) ; both on one Cloud Mac — isolate workspaces.

Local AI: privacy option for Skill workflows

Settings → AI & Skills → Local AI — three presets:

• Embeddings only : all-minilm etc. for Memory Tree ;
• Embeddings + learning : partial background jobs local ;
• Everything local : embedding, summary, heartbeat, learning, subconscious via Ollama.

Local-first mainly Memory and Skill metadata — default chat, vision, web search, OAuth proxy may stay cloud. Ollama Metal on M4 — common compromise — M4 AI dev.

Skill stack на Mac mini M4 Cloud Mac

• Installed Skills + SKILL.md + manifest ;
• Memory Tree (chunks.db + wiki/) ;
• Ollama model caches ;
• 24/7 Auto-fetch needs host awake.

1. Windows/Linux primary : remote macOS for OpenHuman + Skill catalog ;
2. Team shared agent node : fixed SSH/VNC, OPENHUMAN_WORKSPACE on persistent volume ;
3. Same machine as iOS CI : Skill experiments + Xcode on M4 — M4 storage FAQ.

Skill deploy checklist

• Install OpenHuman release, Settings → AI & Skills browse catalog ;
• Install 1–2 Skills, test @skill-name ;
• Connect at least one Auto-fetch source ;
• Backup OPENHUMAN_WORKSPACE (skills dir + wiki/) ;
• Watch v0.55+ release notes for Skill runtime rebuild.

FAQ

Replace ChatGPT GPT? Often better for local auditable workflow + personal memory ; ChatGPT simpler for zero-config cloud multimodal.

Write own Skills? Yes — SKILL_SPEC, manifest, local install or PR to official catalog.

QuickJS still? Removed from main app ; SKILL.md injection primary, runtime rebuilding.

Vs OpenClaw plugins? OpenClaw = multi-channel gateway ; OpenHuman = desktop personal agent + Memory + Skill. Coexist with directory isolation.

Mac required? macOS/Windows/Linux ; 24/7 + Ollama Metal → Cloud Mac common.

Вывод

OpenHuman buzz не от «bigger model», а от «local Skills + Memory Tree + readable wiki» agent infrastructure narrative. Extension logic из cloud GPT Store → installable, reviewable SKILL.md workflows coexisting with your data — runtime beta, direction clear: agent capability boundaries should live on your machine like code. Mac mini M4 Cloud Mac: OpenHuman + Skill catalog + optional Ollama — personal agent engineering path 2026 worth trying.