Get started
For providers
Sell inference on Numes from your own hardware or your upstream API accounts. How you connect, what the platform checks and how you're paid.
Numes is a marketplace: customers send OpenAI-compatible requests, and independent providers serve them. You publish models at prices you choose, the platform routes matching requests to you, and you earn for what you deliver. You never handle customer accounts, API keys or billing.
If you want to call models rather than sell them, start with the Quickstart.
Use an existing Claude Code login
For a self-hosted source, choose Claude Code to use Claude's login on the machine running the agent. Anthropic API uses an API key as before. Install Claude Code 2.1.280 and run claude auth login as the same user that runs the agent. In desktop setups over SSH, do this on the remote machine. Then connect the source, choose models from the live catalogue, review limits and prices, and publish. A Claude Code context variant such as claude-opus-5-5[1m] is listed as claude-opus-5-5-1m and gets the same suggested model name as its base model. Other CLI versions and managed hooks/settings are not qualified; the source check rejects them.
Each active request runs a separate CLI worker, bounded by your endpoint's Concurrent requests setting and your Claude account's limits. Conversation history comes from the customer client. Tools declared by Pi or OpenCode run on the customer's machine; the provider exposes no local files, shell or agent tools. Every model Claude Code reports as thinking publishes reasoning with Claude Code's effort levels. Thinking stays on for each request unless the model is known to allow turning it off, and requested effort levels pass to Claude Code. Thinking counts as output tokens. Cache reads and writes are reported separately; without cache prices they are charged at your prompt price. The agent reads native token usage for billing and leaves credentials in Claude's store. Context/output limits absent from discovery need review before publication.
Claude Code tools support safe integers (from −9,007,199,254,740,991 to 9,007,199,254,740,991) and decimals that its JSON parser can preserve. Define larger identifiers and high-precision values as strings in your tool schemas. Unsupported numeric inputs are rejected; generated tool arguments outside these limits fail before the tool call completes.
Terms
| Term | Meaning |
|---|---|
| Provider | Your public name, shown next to every model you serve. It belongs to a personal or organization account, so teammates can manage it with their own sign-ins. |
| Endpoint | One serving setup: one inference source reached through one connection. The agent stores each endpoint locally as a profile. |
| Source | What actually runs the model: an Ollama server, a vLLM deployment, an upstream API account and so on. |
| Listing | One model on one endpoint, with its capabilities, limits and prices. Every change creates a new revision. |
Provider logo
A provider can show a logo next to its name wherever customers and administrators see it: the provider directory and its detail page, the providers serving each model, and your own provider list. To add one, open your provider page in the dashboard and choose the square next to the provider's name. With a logo set, the same square offers Replace logo and Remove logo. Owners, administrators and provider operators of the owning account can change it.
Upload a still JPEG, PNG or WebP image of up to 10 MB. Numes removes its metadata and stores it as a PNG of at most 512 pixels per side, shown square at 16 pixels in lists and 24 pixels in headings, so a simple square mark reads best. A logo is public as soon as you save it and is not part of provider review.
Deleting a provider or its personal account also removes the logo. Stored image cleanup runs in the background; previously cached copies may remain visible.
Two ways to sell
| Self-hosted | Hosted | |
|---|---|---|
| What runs where | You run the numes agent on your own machine, next to your models or with your upstream API key | The platform runs the source for you with your upstream API key |
| Sources | Ollama, vLLM, llama.cpp, any OpenAI-compatible server, and supported upstream APIs | Supported upstream APIs |
| Setup | Terminal UI, CLI or desktop app | Dashboard only |
Self-hosted is the general path. The agent is a single executable that connects out to the platform over HTTPS, so you open no ports and need no public certificate. See the Provider quickstart.
Hosted needs no machine: you paste an upstream API key in the dashboard and choose models. You pay the upstream as usual; customers pay you. See Hosted sources.
You choose when you create a provider, and you can't change it later. To sell both ways, create one provider of each kind.
How a request reaches you
Your agent keeps an outbound connection to the platform. Nothing connects in to you.
A customer sends a request for a model, such as
openai/gpt-6-luna.Routing picks among live offers that match the customer's constraints and have free capacity. Your offer competes on price, reliability, speed and capacity.
The platform sends the request over your connection. The agent translates it for your source and streams the result back.
The customer is charged your prices for the usage delivered, and the same amount is credited to your pending earnings.
If your endpoint fails before any output reaches the customer, the platform may send the request to another provider, and you earn nothing for it. Once output has been delivered, the request is never switched to another provider. Provider-caused interruptions may be excluded from earnings.
During platform updates, requests that are already running finish normally. Your agent may briefly stop accepting new requests while it reconnects.
What the platform checks
The platform applies these checks to all providers and may re-review any provider, endpoint or offer at any time.
| Check | What it means for you |
|---|---|
| Provider and endpoint review | An administrator approves your public identity and each endpoint's declaration, including its data handling. |
| Model names | Public model names are reviewed by the platform. |
| Price protection | Price changes close to your last reviewed price activate automatically; first listings and large changes are reviewed. See Pricing. |
| Verification | Automated checks exercise every capability you declare, such as streaming, usage reporting, tools, structured output, media and cancellation. A new model or newly declared behavior is checked after review; metadata and price changes reuse the existing result. See Review and lifecycle. |
| Measurements | Time to first token, throughput, errors and capacity feed routing. Repeated failures pause a model until a targeted recovery check passes. See Measurements and quality. |
Each new self-hosted technical contract needs review and verification before going live. Metadata and commercial revisions reuse that pass while current price protection and readiness checks still apply. A hosted offer revision goes live once its name and prices are approved.
Misrepresenting a model, its capabilities, its usage or its data handling leads to suspension and to withholding or reversal of earnings.
How you earn
You set an exact price for each meter of each model: prompt and completion tokens, cache reads and writes, requests, images, audio seconds or hosted tool calls. Customers pay your price for inference. The platform fee is charged when customers buy credits, so it doesn't reduce your earnings.
| Stage | What happens |
|---|---|
| Pending | Your price for the delivered usage is credited to pending earnings. |
| Available | After a holding period, earnings become available. A portion may be held in reserve. |
| Withdrawal | Withdraw available USD to a bank account, or available BTC over Lightning or on-chain. |
USD and BTC earnings are separate balances; they're never added together or converted. See Earnings and payouts.
Supported platforms
The numes agent runs on Linux x64 and ARM64 with glibc, and on macOS with Apple silicon. It has its runtime built in, so the host needs no Node.js, Python or other toolchain. Your inference source can run anywhere the agent can reach, including another machine on your network. See Install and update.
Connections during platform updates
During planned platform updates, accepted requests can finish while the endpoint continues accepting work. All connections share the profile's source capacity limit. This does not protect against a stopped agent or source, or a hard failure during a stream.