Relays and aggregators

This page is for you if you want many models on one key, lower prices than the model makers post, or a way to pay without a card. A relay can give you all three. The price is that one more company sits between your agent and the model, and that company can read everything your agent sends.

The page explains what a relay can see, compares the paid relays this guide checked, and walks through connecting one to your agent. It ends with how to judge a relay before you hand it money or data.

Applies to: Agent Zero, Hermes Agent, OpenClaw, Goose, NanoClaw, QwenPaw, Eigent, Atomic Agent, and OpenManus. Everfern has no relay setting: its only address box is for a local Ollama or LM Studio server and takes no key, though OpenRouter is one of its named providers.

What a relay is

Most agents can talk to any service that speaks the same request format as OpenAI's API. To switch, you change two settings: the address the agent sends requests to, usually called the base URL, and the key it sends along. A relay is a service at a different address that accepts those requests and passes them on to whoever actually runs the model.

An address that speaks OpenAI's format does not mean OpenAI runs it. The relay's own terms and privacy policy apply, not the model maker's.

Relays come in two broad kinds.

• Aggregators sell access to many models and providers through one account and one key. They usually pick the provider for you. NanoGPT, RelaySwitch, and Chutes are the examples on this page.
• Discount relays sell prepaid credit at prices below the model maker's list price. Many resell access to accounts they hold with the model makers. OhMyGPT is the example here.

For comparison, OpenRouter is a widely used aggregator that says it bills at each provider's list price and charges its fee on credit purchases instead, 5.5% for card payments. It works the same way in an agent and carries the same basic question: who else can read your prompts?

What a relay can see

A relay sees everything your agent sends, and an agent sends far more than your typed message. Each step can include the agent's instructions, the conversation so far, files it opened, web pages it read, command output, and notes from its memory. If the agent can reach a folder, whatever it reads there can end up on the relay's servers.

The relay then passes the request to an upstream provider, which sees it too. Most relays choose that provider for you, and some do not say who it is.

A relay also decides which copy of a model answers. NanoGPT's docs say providers can serve a model at different weight precisions, and they offer filters "when you need to avoid lower-precision provider routes for a model". So the default route can be a more compressed copy than another provider selling the same name. RelaySwitch's own daily check marked 20 models "changed" on the day it was checked, meaning what answered under that name was not what answered before. The apiprice directory, which lists discount relays, says a price sticker does not show whether the model you received is the one you named.

Warning: Keep passwords, private code, customer records, medical or legal documents, and anything you would not paste into a public chat away from a relay. That includes files your agent can open on its own.

Your relay key is a target too. Anyone who has it can spend your balance, so store it the way your agent's guide says to store an API key.

The relays this guide checked

Prices and claims below are what each site said on 14 November 2026. "Their claim" means this guide could not check it independently, and nothing here was tested by sending a prompt.

A prepaid balance has one real advantage: it is its own spend limit. When it runs out, the agent stops, so a stuck agent cannot spend more than you loaded.

NanoGPT

NanoGPT's pricing page offers pay as you go at $0 a month, or $12 a month for included use of open-source models. The $12 plan covers 60 million input tokens a week, free output on the included models, and 100 images a day, and it resets every Monday at 00:00 UTC. It does not include most closed models, naming Claude, GPT, Gemini, and Grok, which keep billing your balance.

The page claims text is billed at list price with no markup. Two opt-in charges sit beside that claim: choosing a specific provider adds 5%, and bringing your own key costs 5% of the normal price. Its payment page says crypto "needs no ID or account" and that card payments are less private, because Stripe processes them.

NanoGPT routes each request to a provider by default. Its docs say provider prices and availability can change between when you look and when you send. It runs a token-count check on its most-used models, but says that check does not cover every model.

RelaySwitch

RelaySwitch prices every model per million tokens and takes prepaid top-ups of $10, $25, $50, $100, $250, or $500. You pay by card or in USDT, USDC, BTC, ETH, or SOL. Its no-identity-check page says an email address opens the account, and that request and response bodies are not stored. Those are their claims, and the machine that runs the model still receives your prompt.

Its comparison table showed Claude Opus 5.5 at $2.716 input and $9.70 output per million tokens, against a list price it gave as $4 and $20. A daily "proof" page sends the same test prompt to each model and flags any change in the answer. On the day the research checked it, 325 of 358 models were stable, 20 had changed, and 13 did not answer. That is their own test of their own service, not an outside audit.

Its own pages also disagreed about the size of the catalog, giving three different model counts at once, and the counts moved during the day. Treat any single number as approximate.

Chutes

Chutes says "Pay per token. No subscription, no minimum, no markup." It has no free tier, and it no longer sells monthly plans. It takes more than 25 payment methods through Stripe, including cards, crypto, and bank transfers, plus the TAO token. One row on its live table was Qwen/Qwen3.8-27B-TEE at $0.039 input and $1.25 output per million tokens.

Its privacy page splits models in two. Standard models are often served by third-party platforms, and "A Model Provider's own staff and systems can access content while they process it". Those providers' retention runs from near zero to "generally up to 30 days", which Chutes says it cannot guarantee. Confidential models, marked in the catalog and usually named with TEE, run inside hardware enclaves designed to keep the machine's operator from reading your content.

Note: Chutes says an automatic or fallback route may pick a standard model. If you need the confidential claim on every request, name a confidential model by its exact ID.

Discount relays, with OhMyGPT as the example

Shops in China and elsewhere sell a prepaid key and tell you to change your agent's address, often at a price below the model maker's list. OhMyGPT, run by DogeNet LLC according to its own terms, is a long-running shop with public price pages. Naming it documents those pages; it is not a recommendation.

Its homepage says one key reaches 294 models, and that it picks a supplier and switches if one fails. It says 96 models are below the official price, but that comparison blends cache hits, input, and output in a 7:2:1 ratio. Billing is in credits at 250,000 per dollar, and each line rounds up to a whole credit. Its own example shows a tiny request costing 3 credits against a list price of 4.

Its terms refuse refunds except when the company cannot continue service because of force majeure. They also grant DogeNet LLC a permanent, worldwide, sublicensable copyright license over the prompts you enter and the material generated from them. Its privacy policy reserves the right to store, review, and investigate your messages, and says the upstream may record them as well.

Its cost docs add one quirk: when your balance runs low, it shrinks the reply limit instead of returning an error. Long answers can then stop early, looking just like a normal length cutoff, until you top up. The only sign is a response header that most agents never show you.

The apiprice directory says reselling can conflict with the model makers' own terms. It points to OpenAI's rules against reselling API access and against use outside supported countries, and to Anthropic's rule against resale without approval. When an upstream account is closed, the route dies, and the person holding prepaid credit has no contract with the model maker.

Connect an agent to a relay

The general steps

1. Read the relay's terms, privacy policy, and refund policy, using the checklist in How to judge a relay.
2. Create an account and top up the smallest amount allowed. NanoGPT starts at $1 by card, RelaySwitch at $10, and Chutes has no minimum.
3. Create an API key in the relay's dashboard. Copy it somewhere safe, because many dashboards show it only once.
4. Find the relay's base URL in the table below, and copy a model ID exactly as the relay's model list shows it.
5. In your agent, add a custom OpenAI-compatible provider with that base URL, your key, and the model ID. The per-agent notes below say where.
6. Send a harmless test message, such as "Reply with the word ready."
7. Open the relay's usage or billing page and check the charge and the model name on that request.

What you should see a reply in your agent, and one small charge on the relay's usage page under the model you picked.

Model IDs differ between relays, even for the same model. Chutes uses names like Qwen/Qwen3.8-27B-TEE, while RelaySwitch uses names like claude-sonnet-5-5. Copy the ID from the relay you are using, not from another site.

Per-agent notes

A few details matter for specific agents. Agent Zero runs in Docker, but a relay is on the internet, so its normal address works and you do not need the host.docker.internal trick used for local models. This guide could not confirm the exact name of Agent Zero's address field.

Hermes rejects models whose context window is smaller than 64,000 tokens, so check the relay's listing for the model's context size. For NanoClaw, NanoGPT, RelaySwitch, and OhMyGPT all mention Anthropic-format support, but this guide could not confirm the exact value each one wants in ANTHROPIC_BASE_URL. Check the relay's own Claude Code setup page for that value.

OpenClaw's docs show a custom provider shaped like this, here filled in for NanoGPT. The docs prefer an environment-variable or secret reference over a key typed into the file. This guide could not confirm the exact syntax, so check OpenClaw's secrets page before you save a real key.

{
  models: {
    providers: {
      "nanogpt": {
        baseUrl: "https://api.nano-gpt.com/api/v1",
        apiKey: "YOUR-RELAY-KEY",
        api: "openai-completions",
        models: [{ id: "MODEL-ID-FROM-THE-RELAY" }],
      },
    },
  },
}

OpenManus reads its model settings from config/config.toml. This example uses OhMyGPT's address.

[llm]
model = "MODEL-ID-FROM-THE-RELAY"
base_url = "https://api.ohmygpt.com/v1"
api_key = "YOUR-RELAY-KEY"

What you should see after a restart, the agent lists or accepts the relay's model, and the test message from the general steps gets a reply.

How to judge a relay before you pay

Spend ten minutes on these questions before you top up. A relay that will not answer them in writing is telling you something.

• Who runs it, and where? Look for a named company in the terms and the country whose laws apply. NanoGPT is NanoGPT LLC, a US company under California law. RelaySwitch is Uplink Switch LLC, under Wyoming law. Chutes is Chutes Global Corp., a Delaware corporation. OhMyGPT's terms name DogeNet LLC. Apiprice notes that OpenAI does not list mainland China or Hong Kong as supported countries, which matters for some resold routes.

• What do the terms take? Read the license section. OhMyGPT's terms claim a permanent, sublicensable license over your prompts and outputs, which is far more than a relay needs.

• What is logged, and for how long? The privacy policy should say whether prompts are stored, for how long, and which upstream providers receive them. Chutes gives a range of up to 30 days for standard models, and OhMyGPT reserves the right to store and review.

• Can you get money back? Mostly not.

  • RelaySwitch says top-ups are not refundable and keeps the balance of an account it closes for a breach.
  • Chutes makes crypto payments final, refunds card payments only in limited cases at its discretion, and forfeits any unused balance when your account ends.
  • OhMyGPT refuses refunds in most cases.
  • NanoGPT points to a separate refund policy.

  Credit at RelaySwitch does not expire, but that does not make it refundable.

• Do the prices add up? Compare the per-token price with the model maker's own price page. Watch for blended averages, credit units that are not dollars, and rounding on small requests.

• Does it show which model answered? RelaySwitch's canary, NanoGPT's token-count check, and NanoGPT's precision field are useful signals. They are still the relay checking itself.

• Do its own pages agree? Mismatched catalog sizes or prices across a site's own pages suggest it changes faster than its documentation.

A relay that passes these checks is still a third party with your prompts. The checks lower the risk; they do not remove it.

What this guide does not cover

Some free or cheap routes work by spending other people's money or access. This guide names them only to explain why it skips them.

• gpt4free and similar scrapers. These pass your prompts through services, often other companies' websites, without those companies' permission. The access is unauthorized, and someone else pays for every request.
• Juggling several free accounts. Signing up for many free tiers and rotating between them to dodge the limits breaks the providers' terms. It also means the provider pays for use it never agreed to give you.
• Shared, resold, or leaked keys. A key someone posted online, sold to you, or shared from their account bills a person who did not agree to pay for your use. A leaked key is a stolen key.

None of these is a cheaper version of a relay. Each one makes someone else pay for your agent without their consent.

Trade-offs: when a relay makes sense

A paid aggregator instead of the model maker's API. You send your agent's requests through NanoGPT, RelaySwitch, or Chutes.

• You gain: many models on one key, and payment by crypto or a small prepaid balance.
• You risk: a third party and an unnamed upstream reading your prompts, and a more compressed model answering under a familiar name.
• Makes sense when: the work is low-stakes, such as summarizing public web pages, and the agent cannot reach private folders. It does not make sense for anything you could not paste into a public chat.

A discount relay. You buy prepaid credit below list price from a shop such as OhMyGPT.

• You gain: a lower price per token on some models.
• You risk: a balance stranded when the upstream account closes, no refund, and terms that may claim rights to your prompts.
• Makes sense when: you load only what you can afford to lose and use it for throwaway tasks. For a beginner, the model maker's own API with a spend limit is the better first choice.

A relay for an agent that can read your files. The relay sees whatever the agent opens, not just what you type.

• You gain: an agent that can work on real documents at a lower price.
• You risk: private files leaving your computer for servers whose logging you cannot check.
• Makes sense when: the agent works only in a container or one folder of non-private files, such as public reports to summarize.

Chutes confidential models. You name a model marked as confidential instead of letting Chutes choose.

• You gain: a design meant to keep the machine's operator from reading your prompts.
• You risk: relying on Chutes' description of hardware you cannot inspect, and a smaller choice of models.
• Makes sense when: you want a relay but care about privacy, and you are willing to name the model ID on every request.

Before you choose, run each option past the three questions in Safety versus capability.

Sources

Checked 14 November 2026.

• What a relay is, the risk table, and the discount-relay category: the unofficial routes chapter and the relays notes
• OpenRouter pricing and fees: https://openrouter.ai/pricing and https://openrouter.ai/docs/faq
• NanoGPT pricing: https://nano-gpt.com/pricing
• NanoGPT payment methods: https://nano-gpt.com/faq/payment-methods-and-cryptocurrencies
• NanoGPT provider routing: https://nano-gpt.com/provider
• NanoGPT provider selection and precision: https://docs.nano-gpt.com/api-reference/miscellaneous/provider-selection
• NanoGPT API host and base URL: https://docs.nano-gpt.com/introduction
• RelaySwitch pricing: https://relayswitch.io/pricing
• RelaySwitch no-identity-check page and base URL: https://relayswitch.io/no-kyc
• RelaySwitch daily model check: https://relayswitch.io/proof
• Chutes pricing and payment methods: https://chutes.ai/pricing
• Chutes base URL: https://chutes.ai/integrations
• Chutes privacy: https://chutes.ai/privacy
• Operators, governing law, and refund terms: https://nano-gpt.com/terms, https://relayswitch.io/terms, and https://chutes.ai/terms
• OhMyGPT homepage: https://www.ohmygpt.com/zh-CN
• OhMyGPT costs, credits, low-balance behavior, and base URL: https://www.ohmygpt.com/docs/zh-CN/costs
• OhMyGPT privacy policy: https://www.ohmygpt.com/zh-CN/privacy
• OhMyGPT terms: https://www.ohmygpt.com/zh-CN/tos
• apiprice relay directory methodology: https://apiprice.app/zh/relays/methodology/
• Agent Zero providers, including Other OpenAI compatible: https://github.com/agent0ai/agent-zero/blob/main/conf/model_providers.yaml
• Hermes Agent providers and custom endpoints: https://hermes-agent.nousresearch.com/docs/integrations/providers
• OpenClaw custom providers: https://docs.openclaw.ai/gateway/config-tools/custom-providers
• Goose custom providers: https://goose-docs.ai/docs/getting-started/providers
• NanoClaw providers: https://docs.nanoclaw.dev/extend/providers
• QwenPaw README: https://github.com/agentscope-ai/QwenPaw
• Eigent provider reference: https://www.eigent.ai/docs/provider-reference
• Atomic Agent configuration: https://atomicagent.io/docs/using/configuration/
• Atomic Agent desktop OpenAI-compatible API (custom base URL) provider and llm.providers[] entry: https://atomicagent.io/docs/guides/vllm/
• OpenManus README: https://raw.githubusercontent.com/FoundationAgents/OpenManus/main/README.md
• Everfern README: https://raw.githubusercontent.com/Everfern-AI/Everfern/main/README.md
• Everfern v1.3.0 setup screens (Server URL (optional) for local engines only, fixed cloud list): https://github.com/Everfern-AI/Everfern/blob/v1.3.0/src/app/setup/page.tsx
• gpt4free terms and privacy, for the section on what this guide skips: https://g4f.dev/terms.html and https://g4f.dev/privacy.html

Next

Uncensored models, and why they are a poor fit for an agent, are on Uncensored models.