AnthroClaw
A focused, Claude-native runtime for AI agents that need memory, tools, recovery, and continuity across customer channels. Self-hosted, inspectable, and released under MIT.
Why most “AI agents” fail after the demo
They answer, then forget
A useful response is not an operating workflow. Without persistent state, the next step still depends on a human.
They break outside the happy path
Rigid automations work on predictable data. Every new exception becomes another development ticket.
They flatten the model into an API
Provider-neutral wrappers often discard the capabilities that make Claude useful: sessions, checkpoints, subagents, and skills.
Six architecture choices for production ownership
AnthroClaw is intentionally opinionated. The goal is not maximum abstraction; it is an agent fleet that a business can inspect, operate, and move.
YAML + Markdown + skills
Review the agent in a pull request, edit it with normal tools, and move it between servers.
One relationship, many channels
A conversation can begin in Telegram and continue in WhatsApp without losing context.
Recoverable actions
Return to a known state and rerun from there when an agent takes an unwanted action.
Memory, search, skills, subagents
Tools are readable TypeScript integrations that can be audited, forked, or replaced.
Manage more than one agent
See sessions, active channels, tool use, and memory across specialized roles.
Self-hosted and transparent
Client data stays on controlled infrastructure and the underlying code remains open.
“Most assistant frameworks become provider routers with a chat UI bolted on top. AnthroClaw is intentionally narrower and more opinionated.”
A foundation clients can verify
MIT-licensed source code instead of an opaque vendor dependency.
Agent configuration, knowledge, and skills remain available to your team.