MicroClaw

Summary

MicroClaw is built for the workload that happens in the chat channel, not in a dashboard. The design premise: the agent is a participant in ongoing conversations, it has to remember who it's talking to, and it has to behave consistently when a user moves from Slack to Discord to WhatsApp mid-conversation. All of that requires durable state and a memory model that outlives any single LLM call.

The engine is written in Rust for steady latency and low per-channel overhead — a single MicroClaw process routinely handles thousands of concurrent chat sessions across multiple channels. The provider abstraction layer sits in front of the LLM call, so routing between Claude, GPT, Qwen, or a self-hosted endpoint is a runtime decision based on cost, latency, capability, and channel context.

Architecture

Channel adapters — normalize every chat platform's event model (messages, edits, reactions, threads, DMs) into a uniform event stream.
Session store — durable session state keyed by (channel, user, conversation). Postgres for canonical state; ClickHouse for analytics.
Memory layers — four tiers: short-term (current turn), session (current conversation), user (everything this user has said across channels), channel (shared context in group spaces).
Agent loop — provider-abstract, tool-capable. Tools are declared in a common schema that maps onto Claude's tool protocol, OpenAI's function calling, and the Qwen-Agent format.
Provider router — cost / latency / capability policies. Can fallback across providers on rate-limit or failure.
Security — FrawdBot inline; per-channel rate limiting; allow-listed tool sets per channel.

When to Choose MicroClaw

MicroClaw is the right pick when: the agent lives in one or more chat surfaces as a first-class participant; you need a single agent identity across multiple channels; durable state and memory across sessions is a hard requirement; or you need provider arbitrage at runtime rather than committing to a single LLM vendor. Customer-support agents, community-management bots, and cross-channel executive assistants are the prototypical workloads.

Tradeoffs vs. Other Runtimes

vs. OpenClaw — MicroClaw is chat-native and lighter-weight; OpenClaw is broader on integrations but heavier and Mac-targeted.
vs. NanoClaw — MicroClaw assumes trusted tool execution; NanoClaw sandboxes every tool call. Use NanoClaw when channel audiences are public or tools include arbitrary code execution.
vs. ExoClaw — MicroClaw is self-hostable; ExoClaw is Organized AI's managed cloud equivalent with less operational overhead.

References

[1] Organized AI on GitHub

[2] The Agent Infrastructure Stack — Organized AI

[3] The Observability Architecture — Organized AI