Cut inference costs without dumbing down your agent.
Route summarize, classify, JSON extraction, and short chat to small/nano models via the zerogpu CLI, executed locally by your agent's Bash tools.
ZeroGPU Router is a smart task router for AI agents. It exposes task-specific skills (summarize, classify, extract JSON, and more) that shell out to the local zerogpu CLI, backed by small language models that run for a fraction of the cost of a frontier model. Its also called HeroGPU
Your agent keeps doing the heavy reasoning. The boring stuff gets routed to ZeroGPU.
- OpenClaw: install the
zerogpuCLI pluszerogpu-router(see agents/openclaw/). Skills run locally through your agent's Bash tools. - Claude Code: install the
zerogpuCLI plus the marketplace plugin and you get 29 auto-invoked skills plus a cost-savings readout (see agents/claude/). - Cheap by default: small models for trivial work, frontier model untouched for everything else.
- Per-call savings: every routed task returns model, latency, and a real
savings_usdfigure. - CLI, no infra: everything runs through the local
zerogpuCLI your agent already calls. No servers, MCP endpoints, or infra to stand up.
You need a ZeroGPU API key. Grab it at platform.zerogpu.ai.
1. Install the OpenClaw plugin (zerogpu-router):
openclaw plugins install clawhub:zerogpu-routerThe plugin's skills provision the zerogpu CLI automatically on first use (via a node install spec), so there's no separate global install. You just need a package manager (npm by default) available.
2. Authenticate:
zerogpu loginYou'll be prompted for your API key (zgpu-api-…).
4. Try it:
summarize this account note: Renewal call with Acme Corp went well overall. They
are happy with uptime but frustrated by slow support response times over the last
two months. Their VP hinted at evaluating a competitor if the SLA does not improve
before the December renewal. They also asked about volume pricing for a second
team of about 40 seats. Action items: loop in support leadership and send an
updated enterprise quote by Friday.
The agent runs the summarize skill, which executes zerogpu summarize locally via its Bash tool, and returns a concise, slightly mechanical summary plus a savings line, for example:
Positive Acme Corp renewal call: happy with uptime, frustrated by slow support
the last two months. VP may evaluate a competitor if the SLA does not improve
before the December renewal; also asked about volume pricing for ~40 more seats.
Next: loop in support leadership, send an updated enterprise quote by Friday.
💰 ZeroGPU savings so far: $1.87 (16,320 Claude tokens offloaded)
The Claude Code plugin ships 32 skills: 29 inference skills that Claude auto-invokes when your request matches ("summarize this", "extract the people and companies from this", "classify by sentiment and topic"), plus the manual signin, status, and cost-savings skills. You can also call any skill manually with /zerogpu-router:<name>. Run /zerogpu-router:cost-savings anytime to see how much you've saved by routing trivial work to ZeroGPU.
Grab a ZeroGPU API key at platform.zerogpu.ai, then:
1. Install the zerogpu CLI (the plugin shells out to it):
npm install -g zerogpu-cli2. Authenticate:
zerogpu loginYou'll be prompted for your API key (zgpu-api-…).
3. Install the Claude Code plugin. Start a Claude Code session by running claude in your terminal, then:
/plugin marketplace add zerogpu/zerogpu-router
/plugin install zerogpu-router@zerogpu
/reload-plugins
4. Try it:
Pull every person, company, and email address out of this support ticket:
Hi team, this is Sarah Chen (sarah.chen@northwind-labs.com, +1 415-555-0182).
Our prod database started throwing connection timeouts around 2:14 AM PT last
night. The on-call engineer Marcus Rivera (slack: @mrivera) restarted the
pgbouncer pod but the issue came back within 20 minutes. Billing should go to
our CFO Priya Patel at priya.patel@northwind-labs.com, billing address 1455
Market St, Suite 600, San Francisco, CA 94103. Please call me back at the
number above.
Claude routes to /zerogpu-router:extract-entities automatically and the gliner2.5-base-v1 edge model returns the spans it found for each label you named — people, companies, email addresses — rather than Claude reading them out itself.
The labels are yours to choose, so project-specific identifiers (internal hostnames, IPs, contract numbers, card last-fours) work as well as the standard entity types: name them in the request. For a passage that is not in English, Claude routes to /zerogpu-router:extract-entities-multi instead.
Full walkthrough (prerequisites, every skill documented in detail, troubleshooting): agents/claude/README.md.
ZeroGPU Router exposes thirty-two auto-invoked skills: twenty-nine task routes and three account utilities.
Chat and generation
| Skill | Workload | Model |
|---|---|---|
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 |
Multimodal chat over text, images, and video | 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 |
summarize |
TL;DRs, abstracts, meeting note summaries | llama-3.1-8b-instruct-fast |
summarize-llama-3-2-3b |
Cheaper summaries at 3B parameters | llama-3.2-3b-instruct-fast |
On OpenClaw this skill is named zerogpu-summarize — OpenClaw bundles its own summarize skill (for URLs, videos, and PDFs), and two skills cannot share a name. The transcribe and speak skills carry the same zerogpu- prefix there, for the same reason. Skill names are otherwise identical across both agents.
Classification
| Skill | Workload | 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 and PII
| Skill | Workload | Model |
|---|---|---|
extract-entities |
Extract 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 |
Moderation and embeddings
| Skill | Workload | 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 | Model |
|---|---|---|
transcribe |
Speech to text for voice agents, meetings, subtitles | whisper-tiny |
speak |
Text to speech with expressive delivery and voice cloning | chatterbox-nano |
Account (manual only): signin, status, cost-savings.
Every task skill prints the model's answer as plain text, and occasionally appends a 💰 ZeroGPU savings line.
| Package | Role |
|---|---|
| agents/openclaw/ | OpenClaw: zerogpu-cli + package/plugin id zerogpu-router + CLI-based skills |
| agents/claude/ | Claude Code: zerogpu-cli + marketplace plugin (/plugin install zerogpu-router@zerogpu) |
