ZeroGPU Router
ZeroGPU Router plugin for OpenClaw: route your AI tasks (summarize, classify, extract JSON, speech, chat) to open-weight and small language models via the zerogpu CLI. OpenAI-compatible and pay-as-you-go — cuts agent inference costs without changing your reasoning model.
ZeroGPU Router
Cut your OpenClaw agent's inference costs. Route your AI tasks (summarize, classify, extract JSON, speech, short chat) to open-weight and small language models instead of burning frontier-model tokens. Our APIs are OpenAI-compatible and pay-as-you-go.
#cost-optimization #model-routing #open-weight-models #small-language-models #classification #pii-redaction #summarization #openai-compatible
What it does
Your OpenClaw agent keeps doing the heavy reasoning. Routine tasks get offloaded to ZeroGPU's open-weight and small models, typically 100–1000× cheaper per call.
Our open-weight models are the most cost-effective on the market right now. You'll find models here you won't see anywhere else, and we add roughly one a day.
- 29 task-specific skills:
zerogpu-summarize,classify-iab,extract-json,zerogpu-transcribe, and more - Nothing to host or register: each skill shells out to the local
zerogpuCLI through the agent'sexectool - Unaffected by CLI model updates: each skill names its own model and calls the CLI's model-agnostic
chat_completions,moderations, orembeddingscommand - Savings you can see: every call logs its model, usage, and a real dollar figure
- OpenAI-compatible, pay-as-you-go: no commitments, no idle GPU cost
Quickstart
Get an API key at platform.zerogpu.ai, then:
# 1. Install the CLI the skills shell out to (3.8.0 or newer)
npm install -g zerogpu-cli@latest
# 2. Log in (prompts for your API key)
zerogpu login
# 3. Install the plugin
openclaw plugins install clawhub:zerogpu-router
Every skill shells out to the
zerogpuCLI, so install it globally and run a one-timezerogpu loginbefore using the plugin. The skills need 3.8.0 or newer, the release that added thechat_completions,moderations, andembeddingscommands; on an older CLI they fail witherror: unknown command 'chat_completions'. In sandboxed/Docker agents, make surezerogpuis available inside the container.
Try it
Ask your agent in plain language. It picks the right skill, shells out to the local zerogpu CLI (which sends your text to ZeroGPU's hosted API for inference, instead of using the host model), and replies with the result plus a savings line. The replies below are illustrative: the models are small, so output is concise and a little mechanical.
Your input leaves your machine. These skills transmit the text you give them to ZeroGPU's third-party service. See Data & privacy below before feeding them anything sensitive. Because the agent can pick these skills from plain-language requests, decide up front what you're comfortable routing off-box.
Summarize a meeting recap:
summarize this: Team, quick recap of today's sync. We agreed to push the mobile
checkout redesign to the March release because the payments integration slipped a
sprint. QA flagged two blocker bugs on Android that engineering will prioritize
this week. Marketing still wants the new pricing page live before the end of the
quarter, so we will revisit scope on Thursday. Please update your tickets before
standup tomorrow.
Reply:
mobile checkout redesign pushed to march release, payments integration slipped a
sprint. qa flagged two android blocker bugs, engineering to fix this week.
marketing wants pricing page live by end of quarter, scope revisited thursday.
update tickets before standup.
model: llama-3.1-8b-instruct-fast · 78 tokens in / 41 out
💰 ZeroGPU savings so far: ≈ $2 (18,730 frontier-model tokens offloaded)
Classify a support ticket against your own labels:
classify this support ticket as billing, bug, feature-request, or account:
"I was charged twice for my October subscription, and the second charge is a
different amount than my plan. Please refund the duplicate and explain the
difference."
Reply:
[
{ "label": "billing", "score": 0.94 },
{ "label": "account", "score": 0.29 },
{ "label": "bug", "score": 0.07 },
{ "label": "feature-request", "score": 0.02 }
]
model: deberta-v3-small · 61 tokens in
💰 ZeroGPU savings so far: ≈ $2 (19,180 frontier-model tokens offloaded)
The skills you get
Classification
| Skill | Workload | Backing model |
|---|---|---|
classify-iab | IAB topic classification | zlm-v1-iab-classify-edge |
classify-iab-enriched | IAB categories plus topics, keywords, intent | zlm-v2-iab-classify-edge-enriched |
classify-domain | IAB classification from a bare hostname, no page fetch | zlm-v1-iab-domain-classifier |
classify-zero-shot | Classify against a flat label list | deberta-v3-small |
classify-structured | Multi-axis schema classification | gliner2.5-base-v1 |
classify-decisions | Dynamic-label decisions, intent and routing | gliner2.5-decide-v1 |
Extraction
| Skill | Workload | Backing model |
|---|---|---|
extract-entities | People, places, companies, dates, custom entities | gliner2.5-base-v1 |
extract-entities-multi | The same, for text that is not in English | gliner2.5-multi-v1 |
extract-json | Pull structured fields into grouped JSON | gliner2.5-base-v1 |
extract-signals | Topics, keywords, intent and other contextual signals | zlm-v1-signal-extract |
Generation
| Skill | Workload | Backing model |
|---|---|---|
zerogpu-summarize | TL;DRs, abstracts, meeting summaries | llama-3.1-8b-instruct-fast |
summarize-llama-3-2-3b | Cheaper summaries at 3B parameters | llama-3.2-3b-instruct-fast |
chat | Default chat: long context, multi-step instructions | gpt-oss-120b |
chat-liquid | Fastest, cheapest chat replies | LFM2.5-1.2B-Instruct |
chat-thinking | Short chat replies with a visible reasoning trace | LFM2.5-1.2B-Thinking |
chat-qwen | Multilingual chat, 100+ languages | qwen3-30b-a3b-fp8 |
chat-qwen3-6-35b-a3b | Content extraction and agent workflows, 16K-token context | qwen3.6-35b-a3b |
chat-deepseek-v4-1-flash | 1M-token context with higher-effort reasoning and function calling | deepseek-v4.1-flash |
chat-glm-5-3-flash | Coding and long-horizon agents, cheapest 1M-token context | glm-5.3-flash |
chat-gpt-4-1-mini | 1M-token context, instruction following and tool calling | gpt-4.1-mini |
chat-gpt-5-6-luna | Cost-optimized high-volume chat, 272K-token context | gpt-5.6-luna |
chat-gpt-5-4-nano | High-volume, latency-sensitive routing and extraction, 400K-token context | gpt-5.4-nano |
chat-gpt-6-luna | Most efficient GPT-6, 272K-token context | gpt-6-luna |
decide-tools | Pick the right tool for a request and generate its arguments | zlm-v1-tool-decisions |
Moderation and embeddings
| Skill | Workload | Backing model |
|---|---|---|
moderate | Safety verdict across OpenAI's 13 moderation categories | zlm-v1-moderation-edge |
moderate-llama | Safe/unsafe verdict with the policy categories named | llama-guard-4-12b |
embed | Text to 384-dimensional vectors for search, RAG, dedupe | all-minilm-l6-v2, bge-small-en-v1.5 |
Speech
| Skill | Workload | Backing model |
|---|---|---|
zerogpu-transcribe | Speech to text for voice agents, meetings, subtitles | whisper-tiny |
zerogpu-speak | Text to speech with expressive delivery and voice cloning | chatterbox-nano |
Account
| Skill | Workload |
|---|---|
cost-savings | Cumulative dollars and tokens offloaded to ZeroGPU |
signin | Sign in and persist API key |
status | Show current sign-in status |
Every skill returns { <task fields>, model, usage, savings }. Browse the full catalog with pricing at docs.zerogpu.ai.
Watch your savings
Live dashboard at platform.zerogpu.ai: token usage, latency, per-tool savings, broken down by agent and time range.
Data & privacy
These skills are not local processing. Every content skill (zerogpu-summarize, summarize-llama-3-2-3b, classify-*, extract-*, decide-tools, moderate, moderate-llama, embed, chat, chat-liquid, chat-thinking, chat-qwen, chat-qwen3-6-35b-a3b, chat-deepseek-v4-1-flash, chat-glm-5-3-flash, chat-gpt-4-1-mini, chat-gpt-5-6-luna, chat-gpt-5-4-nano, chat-gpt-6-luna, zerogpu-transcribe, zerogpu-speak) passes the text you supply to the local zerogpu CLI, which transmits it over the network to ZeroGPU's hosted models. The CLI runs locally; the inference does not.
Before using these skills:
- Do not send secrets, credentials, API keys, or regulated data (PHI, cardholder data, etc.) unless you have cleared third-party processing with ZeroGPU for that data.
- Treat inputs the way you'd treat any third-party API call: assume the request may be logged or retained per ZeroGPU's terms and retention policy. Review those at zerogpu.ai for your compliance needs.
Credentials: zerogpu login (the signin skill) writes your API key to a local config file and upserts ZEROGPU_API_KEY into your shell profile, a persistent change to your environment. Revoke keys from the dashboard.
Support
License
MIT. See LICENSE.
