TensorFold for OpenClaw

Run OpenClaw's agent on your own Mac or NVIDIA GPU. OpenClaw starts TensorFold when a request needs it.

TensorFold serves language models from a single Zig binary on Apple Silicon and NVIDIA GPUs, with no Python or MLX to install. This plugin makes it an OpenClaw model provider.

Quick start

brew install ashhart/tensorfold/tensorfold
openclaw plugins install clawhub:openclaw-tensorfold
openclaw onboard --auth-choice tensorfold-local

Setup asks for a model, a context window and whether to stop TensorFold when idle. After that, the next agent turn starts TensorFold on its own.

What you get

• TensorFold starts on demand. OpenClaw launches tensorfold serve on the first request that needs it, and stops it when OpenClaw exits or after an idle time you choose.
• The model's real limits. The provider reads /v1/models, so OpenClaw knows each model's context window, whether it takes images, and the reasoning levels its chat template names, instead of guessing from the name.
• Thinking levels that land. OpenClaw's thinking levels reach the server as reasoning_effort, and off becomes none only when the template can switch thinking off.
• Private by default. The server listens on 127.0.0.1 behind a key kept in a file only you can read.
• Your own server works too. Point the provider at a tensorfold serve you run yourself, on this computer or another.

Requirements

• OpenClaw 2026.11.3 or later
• TensorFold 1.0 or later (checked with 1.2.8): brew install ashhart/tensorfold/tensorfold
• Node 24.16 or later, as OpenClaw itself needs

Let OpenClaw start TensorFold

openclaw plugins install clawhub:openclaw-tensorfold
openclaw onboard --auth-choice tensorfold-local

Setup asks three things: the model (models you have already downloaded are listed first, or give a folder or a Hugging Face id), the context window (65,536 tokens unless you say otherwise), and whether to stop TensorFold when it sits idle. It then starts TensorFold once to read the model's real window and reasoning levels, stops it, and writes OpenClaw's local model service entry under models.providers.tensorfold.

From then on, the first request that needs the model starts tensorfold serve on 127.0.0.1 (a free port from 18180 up), waits for its /health, and sends the request. OpenClaw stops the server when it exits, or after the idle time you chose. The server and OpenClaw share one key, kept in tensorfold/server.keys (mode 0600) under OpenClaw's state folder, so a program needs that key to use the server, and other users on the computer can't read it.

The plugin looks for tensorfold on PATH and in Homebrew's folders. Set TENSORFOLD_BINARY to an absolute path to use another build. For unattended setup:

TENSORFOLD_MODEL=TensorFold/Qwen3.8-27B-MLX-4bit TENSORFOLD_CONTEXT=65536 TENSORFOLD_IDLE_STOP_MINUTES=15 \
  openclaw onboard --non-interactive --accept-risk --mode local --skip-health --auth-choice tensorfold-local

Or connect to a server you run

tensorfold serve /path/to/model --name my-model --host 127.0.0.1 --port 8080 --api-key-file server.keys
openclaw plugins install clawhub:openclaw-tensorfold
export TENSORFOLD_API_KEY=<a key from server.keys>
openclaw onboard --auth-choice tensorfold-api-key

See TensorFold's API key docs for the key file format; a key scoped to inference,read is enough. A server started without keys works too, but OpenClaw still wants a key to register the provider, so give it any value.

Onboarding saves the key and sets the first served model as your default. The plugin uses http://127.0.0.1:8080/v1 (tensorfold serve's default) unless you set TENSORFOLD_BASE_URL or models.providers.tensorfold.baseUrl. A bare address such as http://gpu-box:8080 gets /v1 added. If you list TensorFold models by hand under models.providers.tensorfold.models, the plugin leaves your list alone and skips discovery.

The clawhub: install takes ClawHub's scanned copy of the release. The same package is on npm as npm:openclaw-tensorfold; OpenClaw asks you to confirm an install from outside ClawHub (--force without a prompt).

Thinking levels

Checked

On an Apple M5 Max (Metal) with TensorFold 1.2.8 and OpenClaw 2026.11.3:

Not checked yet: the idle stop inside a long-running Gateway, and a CUDA host.

Development

npm install
npm test

npm test builds dist/ and runs the tests against a fake server and a fake tensorfold binary. To try a build in OpenClaw:

npm pack
openclaw plugins install npm-pack:./openclaw-tensorfold-0.2.2.tgz --force
openclaw plugins inspect tensorfold --runtime --json

License

Apache-2.0, the same license as TensorFold (see LICENSE and NOTICE). Releases 0.1.0 and 0.2.0 were MIT.