MicroClaw

MicroClaw is an agentic AI assistant runtime written in Rust, designed around the principle of "one runtime for all chat surfaces." It drives Telegram, Discord, Slack, Feishu, IRC, Web, and more from a shared Agent Loop architecture, eliminating the need to build separate bots per platform.

On the agentic capabilities side, MicroClaw supports multi-step tool execution (bash, file operations, search, etc.), tool-result reflection, sub-agent delegation with restricted tool sets, and a Plan & Execute pattern that breaks complex tasks into parallel sub-tasks.

The memory system is a standout feature, employing a tiered architecture: L0 Identity (PROFILE) → L1 Essential (high-confidence memories) → L2 Relevance (query-matched memories) → on-demand deep recall. The backend uses SQLite for structured memory storage and a knowledge graph (subject-predicate-object triples), with a background reflector for automatic fact extraction, memory quality gating, and lifecycle management (confidence + soft-archive). Optional sqlite-vec integration enables KNN semantic vector retrieval.

Additional capabilities include cron scheduled tasks (manageable via natural language), an Anthropic Skills-compatible skill system (with sync from external registries like vercel-labs/skills), an MCP protocol client (child-process and streamable-http transports, with memory operations routing through MCP backends and auto-fallback to SQLite), and a local Web control panel built on axum (with Memory Observability).

The project uses a Cargo workspace with modular crates—core, storage, tools, channels, app, observability, and clawhub—achieving separation between channel adapters and core logic through a platform-extensible architecture. It supports Anthropic natively and any OpenAI-compatible API as LLM backends. Current version 0.1.52, in early rapid iteration with core functionality chains complete.