Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

206 — LobsterAI Agent Architecture Analysis


topic: agents type: research status: research-complete last-validated: 2026-05-21 original-query: Evaluate LobsterAI agent architecture patterns transferable to ZAO OS skill system (reconstructed) tier: reference

Status: Research complete Date: March 28, 2026 Goal: Evaluate LobsterAI (NetEase Youdao) for patterns transferable to ZAO OS agent/skill architecture

Key Decisions / Recommendations

Decision Recommendation
Adopt LobsterAI as reference? USE as reference for agent architecture patterns only — not for core features. MIT-licensed, 4,706 stars, actively maintained.
Skill system pattern ADOPT the SKILL.md manifest + skills.config.json ordering/enabling pattern for ZAO OS's planned ElizaOS agent skill system
Memory extraction ADOPT the auto-extraction of user preferences from conversations — useful for ZAO OS's planned AI agent onboarding assistant
Permission gating ADOPT the explicit tool-approval model for any ZAO OS agent that takes actions on behalf of users
IM bridge pattern INVESTIGATE for Farcaster/XMTP bot — LobsterAI's Telegram/Discord gateway lets users control desktop agents from mobile
Music search skill SKIP — scrapes Chinese cloud drives for downloads, not relevant to ZAO OS's legitimate multi-provider streaming player
Desktop app (Electron) SKIP — ZAO OS is a Next.js web app, desktop client architecture is not transferable

What LobsterAI Is

LobsterAI is an all-in-one personal AI agent desktop app built by NetEase Youdao. It runs as an Electron app with 28 built-in skills, executing tasks locally with user permission gating. Tagline: "Your 24/7 all-scenario AI agent that gets work done for you."

  • Stars: 4,706 | Forks: 674 | Open Issues: 579
  • License: MIT
  • Created: February 12, 2026 | Last commit: March 28, 2026 (daily activity)
  • Core engine: Claude Agent SDK (Anthropic) — same AI provider as ZAO OS

Comparison of Options: Agent Architecture Patterns

Pattern LobsterAI ElizaOS (ZAO planned) Custom ZAO Agent
Runtime Electron desktop (local) Node.js server Next.js API routes
AI Engine Claude Agent SDK ElizaOS framework Claude API direct
Skill System SKILL.md manifests, hot-loadable, skill-creator meta-skill Plugin system with character files .claude/skills/ directory
Memory Auto-extracted from conversations, SQLite local Character memory + RAG Supabase + pgvector (planned)
Permission Model Explicit approval per tool invocation Role-based Session-based (iron-session)
IM Integration Telegram, Discord, DingTalk, Lark, NetEase IM Discord, Telegram, Twitter Farcaster casts, XMTP DMs
Storage SQLite (local-only) PostgreSQL or SQLite Supabase PostgreSQL
License MIT MIT N/A

Tech Stack Breakdown

Layer LobsterAI ZAO OS Equivalent
Framework Electron 40 Next.js 16
Frontend React 18 + Redux Toolkit React 19 + React Query
Styling Tailwind CSS 3 Tailwind CSS 4
Build Vite 5 Turbopack
Storage SQLite (sql.js) Supabase PostgreSQL
AI Claude Agent SDK Claude API (planned)
Testing Vitest Vitest

Transferable Patterns for ZAO OS

1. Skill Manifest System

LobsterAI's SKILLs/ directory uses SKILL.md files as declarative manifests — each skill defines its name, description, tools needed, and execution instructions. A skills.config.json controls ordering and enable/disable state. The skill-creator meta-skill lets users author new skills at runtime.

ZAO OS application: The planned ElizaOS agent (see research/205-openclaw-paperclip-elizaos-deployment-plan/) could adopt this manifest pattern for its skill/plugin system. Currently ZAO OS has .claude/skills/ for Claude Code skills — a similar pattern for runtime agent skills would be natural.

2. Persistent Memory Extraction

LobsterAI auto-extracts user preferences and personal facts from conversations, stores them in SQLite, and injects them into future prompts with configurable limits. This runs passively during all conversations.

ZAO OS application: For the planned AI onboarding assistant (research/200-community-os-ai-agents-platform-vision/), auto-learning member music preferences, governance participation patterns, and communication style would personalize the experience. Store in Supabase with src/lib/db/supabase.ts.

3. Cowork Session Architecture

LobsterAI's "Cowork mode" creates autonomous working sessions where the agent executes a sequence of tool calls with user approval gates between each step. This is different from chat — it's task execution with checkpoints.

ZAO OS application: A "cowork" pattern for governance proposal drafting — the agent researches, drafts, formats, and submits proposals with member approval at each step. Integrate with src/components/governance/ workflow.

4. IM Remote Control Gateway

LobsterAI's IM bridges (Telegram, Discord, DingTalk, Lark) let users send commands from their phone to the desktop agent. The agent executes locally and reports back via the same IM channel.

ZAO OS application: A Farcaster cast-triggered bot or XMTP DM bot could let ZAO members trigger agent actions (e.g., "play [song]", "check my respect score", "draft a proposal about [topic]") from any Farcaster client. Wire through src/lib/farcaster/neynar.ts webhook handler.

ZAO OS Integration

Relevant files:

  • src/lib/farcaster/neynar.ts — webhook handler for potential Farcaster bot commands
  • src/components/governance/ — governance proposal workflow that could use cowork-style agent assistance
  • research/200-community-os-ai-agents-platform-vision/ — AI agents platform vision
  • research/205-openclaw-paperclip-elizaos-deployment-plan/ — ElizaOS deployment plan
  • scripts/openclaw-setup-github.sh — OpenClaw setup (LobsterAI also integrates OpenClaw)
  • community.config.ts — admin FIDs that would control bot permissions

Integration priority: Low-medium. LobsterAI is a reference architecture, not a direct integration. Extract patterns for the post-MVP AI agent layer.

28 Built-in Skills (for reference)

Web search, document generation (docx/xlsx/pptx/pdf), video generation (Remotion), web automation (Playwright), canvas design, frontend prototyping, game development, scheduled tasks (cron), email (IMAP/SMTP), stock analysis, music search, film search, content planning, article writing, weather, and more.

The scheduled tasks skill (cron expressions for daily digests, periodic reports) is directly relevant to ZAO OS's planned automated governance summaries and weekly Respect reports.

What's NOT Useful for ZAO OS

  • Desktop-only architecture — ZAO OS is a web app, Electron patterns don't transfer
  • Single-user model — LobsterAI is personal assistant, ZAO OS is community platform for 100+ members
  • Chinese cloud drive music search — scrapes Quark/Baidu/Aliyun for downloads, irrelevant to legitimate streaming
  • Redux state management — ZAO OS uses React Query, different paradigm
  • SQLite local storage — ZAO OS uses Supabase PostgreSQL with RLS

Sources