Docs

Marketplace

Tokmine is also a marketplace of private AI providers. Workspaces can sell idle capacity from their machines and buy capacity when their own machines are missing or busy. Live prices are on the marketplace page.

Requests run on third-party machines

When you buy, your prompts and the model output pass through a provider's computer. Tokmine does not store request or response bodies, but it cannot control what a provider's machine does. Do not send secrets or regulated data, and treat output like any third-party API.

Selling (providers)

Install the CLI, run tokmine and tokmine login so your machines belong to a workspace.
In the app open Marketplace, Sell. Set a minimum price per 1M tokens (input and output, with a default and per-model overrides), a weekly schedule with a timezone, which machines are shared, and the maximum concurrent marketplace requests per machine.
Accept the marketplace terms. A valid card on file is required, and checked on the server.

Your own workload always has priority. Outside your schedule, or when selling is off, no marketplace request is routed to your machines. tokmine status shows whether you are selling right now and why. Selling is configured in the app, not in the CLI.

Buying

Set a maximum price per 1M input and output tokens and an optional monthly budget under Marketplace, Buy.
Marketplace use is off by default. Turn it on for the default endpoint, or per endpoint.
It is a fallback: a request goes to the marketplace only if your own machines cannot serve it, because none has the model online or every candidate is saturated or unhealthy. It is never used when your own pool can serve it.
You can require privacy-first providers, or only prefer them (the default).
Providers asking more than your maximum are excluded. When the monthly budget is reached the request is refused with 402 marketplace_budget_exceeded.
You can never buy from your own workspace, and anonymous sessions can neither buy nor sell.

Prices and the fee

Prices are USD per 1M tokens, separate for input and output, per model. A provider's asking price is what the buyer pays. Tokmine keeps 10% of every sale and the provider receives 90%. The charge for a request is computed from the tokens reported by the response:

charge = tokensIn * askIn / 1,000,000 + tokensOut * askOut / 1,000,000

Token counts come from the response usage and are sanity-checked against the bytes sent and received. If a response has no usage, tokens are estimated from size and marked as estimated. Amounts are tracked as whole micro-dollars, never floating point.

Billing

Usage is metered as it happens and settled monthly through invoicing. Buyers get an invoice line and providers a payout credit, with a monthly statement for each workspace. There is no prepaid balance. Both selling and buying need a valid card on file.

The public price list

The marketplace page shows the lowest, median and highest asking price per model, the number of providers and the share that are privacy-first. It is anonymised: it never includes workspace or machine identifiers, and a model with too few providers shows only limited supply. The same data is served as JSON by GET /v1/marketplace/prices on the agent endpoint, values in micro-USD per 1M tokens.

Privacy and anonymity

Buyers and providers are never named to each other. Providers are described as privacy-first when the CLI found no prompt logging, see privacy checks for what that does and does not mean. Machines where prompt logging is detected are never offered.

Limits worth knowing

Providers can be slower, smaller or less reliable than a hosted API, and output can be wrong. Validate what you rely on.
Providers that report implausible token counts are clamped and, if it repeats, removed from routing.
Rules on resale, content and abuse are in the acceptable use policy.