Skywork ships SkyClaw v1.0, an agent model tuned for OpenClaw, Hermes, and Nanobot

Free access is available via APIFree, plus a faster lite variant and an OpenAI-compatible agent chat API with thinking mode, tool calls, and long-context stability.

By Ryan Merket · 2026-05-19T14:56:41.939Z

Why it matters

Agent stacks live or die on tool reliability and multi-step stability. A model purpose-built for OpenClaw, Hermes, and Nanobot could cut integration pain and make agents production-ready faster.

A digitally rendered, intricate 'claw' or multi-branching data structure, composed of glowing lines and abstract forms, actively interfacing with a complex, ethereal network of information. (Photorealistic 3D render, meticulously textured a

Skywork launched SkyClaw v1.0, a production-focused agent model built for popular agent environments OpenClaw, the Hermes agent framework, and Nanobot, in a thread on X. The team is also shipping a lighter, faster variant.

What shipped

SkyClaw v1.0: a high-performance, agentic LLM optimized for tool use, long-horizon planning, structured output, and stable behavior in long contexts, according to the model page.
SkyClaw v1.0-lite: a speed-focused variant the team says trades a bit of quality for lower latency on complex agent tasks.
OpenAI-compatible Agent chat API: exposed at POST /agent/v1/chat/completions on APIFree, with streaming, function calls, and configurable thinking mode.

Skywork says SkyClaw was trained and tuned directly against the agent stacks developers are deploying today. "Built on a self-constructed OpenClaw environment with high-quality tools and synthesized tasks derived from real user patterns," the team wrote in a thread on X.

How it works (for builders)

SkyClaw exposes an OpenAI-style interface on APIFree, so most client libraries should work out of the box:

Streaming responses for real-time agent UX.
Function/tool calling for external integrations.
Multiple "thinking" modes you can toggle via chat_template_kwargs.
Context caching to cut cost and latency on long conversations.
MCP support to plug in external MCP tools and data sources.

The docs on APIFree include a curl example, max token settings (up to 32,768), and key setup via the API keys page.

Benchmarks (as reported by Skywork)

In internal and synthetic evaluations aimed at agent behavior, Skywork reports that both v1.0 and v1.0-lite outperform Minimax 2.7, DeepSeek V4 Flash, and Qwen across PinchBench, Claw-Eval (with ^3 stability), and a Skywork-Claw-Bench suite, per the X thread. The company has not yet published external replications, and the benchmark image is shown on the APIFree model page.

How to try it

Skywork is offering access through APIFree, with a free tier available. The team pointed developers to sign up for APIFree in the thread and to call the model as skywork-ai/skyclaw-v1 via the API.

The bet

Agent frameworks are proliferating, and many general-purpose LLMs still stumble on multi-step plans, tool discipline, and long-context stability. Skywork is betting that a model trained and evaluated inside the exact agent environments teams use day to day will reduce integration friction and make agents feel steadier under production load. If the reported gains hold up in public evals and real apps, SkyClaw could become a practical drop-in for OpenClaw, Hermes, and Nanobot deployments that want stronger tool use without a ground-up stack change.