mold

Appearance

OpenClaw ​

OpenClaw is an agentic coding assistant and runtime that can load workspace-specific skills. Mold's CLI-native interface makes it a natural agent tool: commands remain inspectable and reproducible outside the agent, while the same local or remote engine is also available through MCP and REST/SSE.

How It Fits Together ​

The usual setup is:

  1. Run mold serve on the machine with the GPU.
  2. Pair the machine OpenClaw runs on with that host (mold host add).
  3. Install the mold skill into the workspace so agents can call the CLI or API.

That keeps image generation on the GPU host while the rest of your agent work happens from a laptop, devbox, or remote shell.

Install the Skill ​

Install the embedded skill directly into OpenClaw's user-wide path:

bash
mold skill install openclaw

For a project-local workspace, run mold skill install openclaw --project from its root or pass --dir <PATH>. Mold writes .agents/skills/mold/SKILL.md for project installs and ~/.openclaw/skills/mold/SKILL.md user-wide. The OpenClaw render includes its {baseDir} path convention and runtime metadata, plus a concise router, operational safety guidance, tested examples, and one shared prompting guide plus one base file under references/prompting/families/ for each manifest family. H3, Wan, and LTX-2 add only their necessary task leaves. This keeps LTX-2.x and MiniMax H3 speech/audio grammar available without loading every family's instructions at once.

Configure the Connection ​

Pairing is done by the user, not the agent. In their own terminal, the user runs mold host invite on the GPU host and mold host add on the OpenClaw machine, then pastes the link at the prompt. An agent never pastes pairing links or API keys; it only uses the saved host by name:

bash
mold host list                      # saved hosts and the default
mold list --host workstation
mold run flux-schnell:q8 "a cinematic product shot" --host workstation

mold host use workstation makes it the default so --host can be omitted; MOLD_HOST=workstation also works where an environment variable fits better. If a command reports "No secure pairing for this host", ask the user to pair it.

Notes ​

  • If the server is remote, HF_TOKEN must be set on the server side for gated model pulls.
  • mold ps --host workstation is the quickest way to confirm the OpenClaw environment can reach the server.
  • The same setup works well for agent workflows that want prompt expansion, gallery generation, or remote batch jobs.