DEV Community

ZeroGPU
ZeroGPU

Posted on

Introducing the ZeroGPU Router plugin for OpenClaw

#ai #openweight #openclaw #slm

Your OpenClaw agent has direct access to your shell and your files, and it runs on a frontier model. That's the right setup for the reasoning parts of a job. It's an expensive way to redact an email address or tag a support ticket.

The zerogpu-router plugin, now on ClawHub, hands that work to specialized small and nano models running across ZeroGPU's edge network. Your agent keeps its reasoning model. The rote steps run somewhere cheaper.

The problem

Agentic runtimes multiply the number of small NLP calls in a workflow. Read the thread, pull the fields, tag the message, scrub the addresses, condense the result. Each of those turns is billed at frontier prices and carries frontier latency, and each one is a well-defined task with a purpose-built model already available for it.

Frontier models for reasoning. ZeroGPU for repeatable execution.

What the plugin does

Install from ClawHub and every ZeroGPU CLI command becomes an OpenClaw skill your agent can read and invoke on its own. Describe the task in plain language, and the agent picks the matching skill, runs the zerogpu CLI locally, and returns a structured result with the model name and token usage attached.

The skills shell out through OpenClaw's built-in Bash tools, so there is no server to host and no API to register. User text is passed to the CLI verbatim through a heredoc, so quotes, newlines, and shell metacharacters in your input are handled safely.

14 skills, one plugin

Eleven inference skills auto-invoke when the agent detects a matching request. Three account skills are manual-only.

Chat and summarization

  • chat — one-line replies that don't need host-level reasoning
  • chat-thinking — short logic and math with a visible reasoning trace
  • summarize — reports, ticket threads, transcripts, and docs

Classification

  • classify-iab — IAB taxonomy
  • classify-iab-enriched — categories plus topics, keywords, and inferred intent
  • classify-zero-shot — against your own flat label list
  • classify-structured — multi-axis schema classification

Extraction

  • extract-entities — custom-label NER with confidence scores
  • extract-json — pulls named fields into a schema-defined object

PII

  • extract-pii — returns PII entities grouped by category
  • redact-pii — masks each span in-line

Account and savings

  • signin, status, and cost-savings — triggered manually

Keeping raw PII out of your host model's context

redact-pii runs on gliner-multi-pii-v1 and replaces each detected span with a [LABEL] placeholder before the text goes anywhere else:

Input:  Email John Smith at john@acme.com about invoice 12345.
Output: Email [PERSON] at [EMAIL] about invoice 12345.
Enter fullscreen mode Exit fullscreen mode

The small model does the masking, and only the scrubbed text comes back. The invoice number survives, because only spans the model recognizes as PII are replaced. Pair it with extract-pii when you also want an audit log of what was masked.

Watch it run live inside an OpenClaw agent →

Seeing what you save

Replies surface a savings line periodically:

ZeroGPU savings so far: ≈ $2 (18,730 frontier-model tokens offloaded)
Enter fullscreen mode Exit fullscreen mode

Token counts come from the API. The dollar figure estimates what the same work would have cost on your host model, with a default baseline of claude-opus-4-8 that you can override with the ZEROGPU_SAVINGS_MODEL environment variable. Run the cost-savings skill any time for the cumulative total.

A worked example: the PII-safe inbox summarizer

Save a triage workflow to your assistant once, then forward it any message. It redacts the personal data, summarizes what's left, and labels the result, all on ZeroGPU's nano models. Full walkthrough in the cookbook.

Getting started

Install the CLI the skills shell out to:

npm install -g zerogpu-cli
Enter fullscreen mode Exit fullscreen mode

Log in with your API key from the dashboard:

zerogpu login
Enter fullscreen mode Exit fullscreen mode

Install the plugin from ClawHub:

openclaw plugins install clawhub:zerogpu-router
Enter fullscreen mode Exit fullscreen mode

You'll need OpenClaw plugin API 2026.4.0 or newer, Node 20 or newer, and a ZeroGPU API key that starts with zgpu-api-. Pin a release with clawhub:zerogpu-router@2.1.2. For sandboxed or Docker-based agents, make sure the zerogpu binary exists inside the container the agent actually runs in.

Top comments (0)