Key points
- MiniClaw (v0.5.0 "The Nervous System", by 8421bit): a ~2,700-line TypeScript MCP plugin with a micro-kernel, parasitic architecture — it plugs into hosts like Claude Code and Cursor, runs via
npxwith zero installation, and focuses on memory + context management rather than its own reasoning engine. - myclaw.net (by linkerlin): a 5,000+ line C# / .NET 9 standalone runtime built on AgentScope.NET and Uno Platform with a modular layered architecture (Channels → Gateway → Agent → Core → Infrastructure), including a ReAct reasoning loop, message bus, WebUI over WebSocket, WinUI desktop GUI, and a CLI.
- Memory distillation: triggers when entries > 20, log tokens > 40% of budget, oldest entry > 8 hours, or log > 8KB
- Memory archiving (
miniclaw_archive) - Entity graph (6 entity types)
- ACE (Adaptive Context Engine): time-aware modes (Morning 06–09 briefing, Work, Break, Evening reflection, Night minimal), session continuation detection, Git/tech-stack workspace awareness, entity surfacing. myclaw.net only loads static templates.
- MiniClaw offers
miniclaw_update/note/read/search/entity/exec/status; skills may include executable scripts. myclaw.net maps several memory operations but lacks entity management and command execution. - Scheduling: MiniClaw uses node-cron (30-min heartbeat only); myclaw.net wins with Quartz.NET supporting full Cron / every / at expressions and scheduled message delivery.
- UI: myclaw.net wins decisively — Uno Platform WinUI, WebSocket WebUI, and a full CLI; MiniClaw depends entirely on its host.
- MiniClaw executes commands inside a strict whitelist (
git,ls,cat,grep,npm,python, etc.;rm/sudo/chown/mv/ddforbidden), with 10s timeout, 1MB output truncation, and working-directory limits. myclaw.net has no command execution. - Both store data locally (
~/.miniclaw/,~/.myclaw/) with no cloud upload. - MiniClaw GitHub: https://github.com/8421bit/miniclaw
- myclaw.net GitHub: https://github.com/linkerlin/myclaw.net
- OpenClaw: https://github.com/openclaw/openclaw
- MCP: https://modelcontextprotocol.io/
- AgentScope.NET: https://github.com/agentscope/agentscope
Architecture comparison
| Dimension | MiniClaw | myclaw.net |
|---|---|---|
| Architecture | Micro-kernel | Modular layers |
| Runtime | Node.js + host editor | .NET (standalone) |
| Protocol | MCP | AgentScope.NET + own Gateway |
| Deployment | npx one-click | dotnet build + CLI |
MiniClaw's three layers: Interface (MCP server, tools, heartbeat), Kernel (ACE engine, memory, skill sandbox), and DNA (AGENTS.md / SOUL.md / USER.md templates).
Memory and context
Both share a dual-layer memory design (daily log + MEMORY.md, append-mode writes). However, MiniClaw wins on:
Tools, skills, and scheduling
DNA/template system
MiniClaw fully implements OpenClaw's 9-file DNA (AGENTS.md, SOUL.md, IDENTITY.md, USER.md, TOOLS.md, MEMORY.md, HEARTBEAT.md, BOOTSTRAP.md, SUBAGENT.md); myclaw.net implements 4 core files. MiniClaw also has an automatic evolution protocol: user signals (e.g., "I like...", "be more casual", "remember this") map to target files and tools without manual intervention.
Security
Recommendations
1. Beginners / individuals: start with MiniClaw — one npx command to experience an AI agent with memory and self-evolution inside your editor.
2. Standalone product needs: choose myclaw.net for its GUI, multi-channel support (Telegram/WebUI), and complex cron scheduling.
3. Combine: MiniClaw as an in-editor daily assistant, myclaw.net as a standalone service for scheduled tasks.