Definition: An AI claw is a persistent autonomous agent that runs continuously in its own sandbox, maintains structured long-term memory, and acts on the user's behalf via tools and messaging interfaces. The term was coined by Andrej Karpathy in March 2026 to distinguish always-on agents from session-based agents. Open-source implementations include OpenClaw (~400K LOC, fully featured) and NanoClaw (~4K LOC, minimal and auditable).
Where The Name Comes From
In a March 2026 demo, Karpathy showed Dobby — a personal smart-home claw he built. Dobby discovered devices on his network via IP scan, reverse-engineered their APIs, and exposed everything through WhatsApp. No app screens, just natural-language messages. The "claw" name stuck because the agent didn't pass tasks back and forth like a session agent — it held on to context, persistently.
Read the deep dive: What Are AI Claws? Persistent Autonomous Agents.
OpenClaw vs NanoClaw
Both are open-source. They sit on opposite ends of the auditability vs feature-richness tradeoff.
| Property | OpenClaw | NanoClaw |
|---|---|---|
| Codebase size | ~400K lines | ~4K lines core engine |
| Feature richness | High (soul docs, episodic memory, WhatsApp portal, experimentation framework, defaults) | Minimal (containerization, full audit, you wire the rest) |
| Auditability | Hard to read end-to-end | Fits in one human or AI-agent context window |
| Best for | Teams that want a working claw on day one | Teams that want to understand every line |
| Security burden | High — you run it, you secure it | Lower (smaller attack surface) but still self-hosted |
| Update cadence | Active community | Slower, deliberate |
Both expect you to host them. Both expect you to be the security engineer.
Where Taskade Fits
Taskade is the managed alternative. The Workspace DNA loop — Memory + Intelligence + Execution — is functionally a claw architecture, but Taskade runs it for you:
- You don't host anything (no servers, no containers, no SSL certificate management)
- You don't manage memory storage (Projects are the memory layer)
- You don't manage tool registries (22+ built-in tools, plus 100+ integrations)
- You don't write the audit trail (every action is logged with attribution)
- You don't set up SSO / OIDC (GenesisAuth, v6.144+)
- You don't build the chat UI (it's built into the workspace)
The trade is: less control over the runtime, far less time spent being a security engineer. For teams that want to ship business outcomes rather than operate infrastructure, the managed path wins.
When To Build Your Own Claw
Pick OpenClaw / NanoClaw / your own claw stack when:
- You have a team of engineers who want to operate the runtime
- You need air-gapped or self-hosted-only deployment (compliance)
- Your use case is specialized enough that a workspace tool would be wrong (Karpathy's smart-home Dobby fits here)
- You explicitly want to read every line of code that runs on your behalf
When To Use Taskade Instead
Pick Taskade when:
- You're a team or non-technical user, not a runtime operator
- You need agents and a workspace and automations and a public app surface — and you'd rather not stitch them
- You want compliance and audit on day one, not as a backlog item
- You want to publish your agents to other users (Community Gallery, public agents)
- You want multi-agent collaboration without writing the orchestration
The "Skill Issue" Argument
Karpathy's broader point in 2026 is that the bottleneck for autonomous AI isn't model capability — it's workflow design. Most teams that "fail at autonomous agents" haven't found a way to string together what's already available. Configuration, instructions, guardrails, decomposition. Whether you ship on OpenClaw, NanoClaw, or Taskade, the discipline of agentic engineering matters more than which runtime you picked.
Taskade reduces the surface area of the discipline. You still need to write good agent instructions, define clear triggers, choose the right tools. But you don't also need to debug a Kubernetes operator at 2am.