MicroClaw

⚡ Local-runtime-first AI assistant platform with host-controlled tools, browser operations, planning, and strong operator ergonomics.

Website · 中文 · Discord

---

🧬 What This Project Is

MicroClaw is built on top of NanoClaw.

Think of it as a powered-up evolution:

• same practical foundation
• stronger execution architecture
• better runtime routing
• deeper browser and tool workflows
• improved operational control

NanoClaw is still the foundation. MicroClaw is the upgraded system we shaped for higher real-world completion.

---

🚀 Why MicroClaw Is More Powerful Than NanoClaw

1) Better runtime intelligence

• Profile-driven runtime selection
• Capability-aware route decisions
• Safer fallback behavior for local/compatible endpoints

2) Stronger execution control

• Host-governed privileged actions
• Cleaner container-runner contracts
• Better execution backend boundaries

3) Real browser operations, not just “search”

• Managed browser sessions
• Snapshot + ref interaction model
• Action-level control and policy shaping

4) Better continuity and memory outcomes

• Structured context building
• Retrieval-aware memory shaping
• Explicit memory tools for long workflows

5) Operator-grade workflow

• Command-center style CLI
• Setup and diagnostics flow
• Better observability and operational ergonomics

---

🧩 Exactly What Changed From NanoClaw

Here is the direct technical delta.

Area	NanoClaw Base	MicroClaw Upgrade
Runtime	Basic profile compatibility	Expanded runtime manager, capability routing, safer fallback paths
Tooling	Core tools	Richer tool registry, better tool exposure rules per route
Browser	Limited interaction pattern	Host-managed browser operator stack with structured action flow
Context	General continuity	Layered context pipeline with stronger memory relevance behavior
CLI/Ops	Core setup	Command-center flow for onboarding, health, debug, and control
Execution	Core container loop	Cleaner host/runner split and execution backend wiring
Channel Surface	Foundation channels	Expanded channel handling plus integration scaffolding

---

🛠 Build Story (How We Made It)

This is the chapter that matters most.

Phase 1: Small assistant core

We started with a minimal orchestrator around:

• message routing
• SQLite state
• isolated execution

Goal: keep the code understandable and safe.

Phase 2: Local-runtime foundation

We evolved into local-runtime-first architecture with:

• runtime profiles
• local endpoint support
• capability probing
• fallback logic

Phase 3: Command center operations

As complexity grew, operations had to get easier.

We added CLI-centered flows for:

• onboarding
• auth/runtime management
• diagnostics
• logs and debugging

Phase 4: Adaptive web layer

Web access was made adaptive, not always-on.

The assistant uses it only when a task truly requires current information or verification.

Phase 5: Browser operator foundation

This was a major leap.

We added host-managed browser operations with:

• centralized manager
• session ownership
• action boundaries
• snapshot/ref workflow

Phase 6: Planner-critic integration

For complex tasks, one-pass responses were not enough.

Planner/critic tooling was introduced for multi-step reliability.

Phase 7: Memory and continuity upgrades

We improved how memory is collected, ranked, and reused so long-running tasks stay accurate.

---

🧠 System Flow

User message
  -> continuity/context assembly
  -> runtime profile selection
  -> capability route selection
  -> tool exposure for that route
  -> model tool loop and/or response
  -> host-managed execution for privileged actions
  -> final response

Capability routes:

• plain_response
• web_lookup
• browser_operation
• deny_or_escalate

---

🔧 Tool Families

Web:

• web_search
• web_fetch
• web_open_url
• web_extract_text
• web_get_links
• web_close

Browser:

• browser_open_url
• browser_snapshot
• browser_click
• browser_type
• browser_select
• browser_extract_text
• browser_screenshot
• browser_tabs
• browser_close

Meta:

• create_plan
• critique_response

Memory:

• remember_this
• memory_search

---

📁 Repository Guide

• src/index.ts: host orchestrator entrypoint
• src/runtime/: runtime selection and capability routing
• src/browser/: browser manager and IPC pathway
• src/context/: continuity and memory assembly
• src/db.ts: persistence
• src/cli/: command-center interface
• container/agent-runner/: runtime adapter and tool bridge

---

⚙️ Quick Start

Requirements

• Node.js 20+
• A local/remote OpenAI-compatible endpoint (LM Studio, Ollama, etc.)
• Optional channel credentials for extra integrations

Install

git clone https://github.com/rithwik1510/MicroClaw.git
cd MicroClaw
npm install
npm --prefix container/agent-runner install

First-Time Setup

npm run dev -- onboard

CLI

Primary command:

microclaw

Compatibility alias:

nanoclaw

Common commands:

microclaw status
microclaw doctor
microclaw models list
microclaw debug
microclaw logs --lines 200

---

📚 More Reading

• docs/SPEC.md
• docs/SECURITY.md
• overall plan.md
• features we need to add.md

---

📄 License

MIT