# Tokex > An exchange for LLM credits. Buy credit for a model, spend it through one OpenAI-compatible API, and trade it for another model's credit in open pools. Tokex issues its own credit per model (for example gpt-5, claude-opus-5, llama-70b). It does not import balances held at providers, and API credit is not a ChatGPT or Claude subscription. Credits are denominated per provider and model; "1 credit" alone means nothing. Status: early access, opening by invitation from the waitlist (https://tokex.dev/#access). Every balance and price on this site is simulated demo credit, and nothing is charged. ## What it does - Exchange: every pair is a constant-product pool. Fees are 0.30% to the pool and 0.05% to Tokex, per leg. Pairs without a pool route through gpt-5, the hub. - Router: POST /v1/chat/completions accepts OpenAI's messages format. The worst case (prompt plus max output) is held before the call; what the answer did not use returns. A failed provider call returns the whole hold. Mutating requests require an Idempotency-Key header. - Auto: model "auto" or "auto:" (fast, balanced, frontier) prices every candidate through the pools and runs the cheapest one the account's credit covers. - Budgets: caps refuse (402 budget_exceeded) before any provider sees the request. - Kill switch: POST /v1/panic revokes every other key and freezes money movement (423 account_frozen); reads stay open. - MCP: a stdio server whose tools include tokex_ensure_credit, which trades for exactly the shortfall an agent needs, and keyed swaps and completions that never run twice. - Index: what a standard job (1M tokens in, 1M out) costs on the cheapest listed model, priced by the pools. GET /v1/index and /v1/index/history take no key. - Fees: 5% load fee on card purchases, shown in the quote; swaps as above; spending burns credit at the model's listed rate. Redemption is built and switched off. ## Pages - [Home](https://tokex.dev/): Buy credit for a model, spend it through one API, and trade it for another model when you need to. Tokex is an exchange for LLM credits. - [Exchange](https://tokex.dev/exchange): Trade one model’s credit for another’s. Every pair is a constant-product pool, and the impact and fees show before you confirm. - [Agents](https://tokex.dev/agents): Point any OpenAI client at Tokex. Autopilot runs the cheapest model your credit covers, and caps refuse a call before a provider sees it. - [Prices](https://tokex.dev/prices): The Tokex Index: what a million tokens in and out cost on the cheapest listed model, priced by the pools. Public, no login. - [Fees](https://tokex.dev/fees): Every fee Tokex charges: 5% to buy credit by card, 0.30% to the pool and 0.05% to Tokex per swap, and the model’s rate when you spend. - [FAQ](https://tokex.dev/faq): Straight answers about Tokex: what it is, whether the money is real, how prices are set, what it costs, and what happens to your prompts. - [Terminal](https://tokex.dev/terminal): The demo market on one screen: the Index and every credit charted live in candles or a line, with the pools and the tape beside them. - [Terms](https://tokex.dev/terms): The Tokex terms of service, a draft for counsel written against how the product actually works. Not yet in force. - [Privacy](https://tokex.dev/privacy): What Tokex collects, what leaves it, and what it never does: no cookies, no analytics, no third-party scripts. A draft for counsel, not yet in force. ## Docs - [Tokex Docs](https://tokex.dev/docs): Tokex is an exchange for LLM credits: buy credit for a model, spend it through one API, and trade it for another model’s credit in open pools. - [Quickstart](https://tokex.dev/docs/quickstart): Mint a key, point any OpenAI client at Tokex, and set a cap. Your first call runs on the cheapest model your credit covers. - [Credits](https://tokex.dev/docs/credits): Credit is issued per model. You buy it by card, spend it through the API, trade it in the pools and, once it is switched on, redeem it. - [Pools and pricing](https://tokex.dev/docs/pools): Every pair is a constant-product pool. A credit’s price comes from the pools, not from a price list, and every quote states its impact and fees. - [The Tokex Index](https://tokex.dev/docs/index): What a standard job costs on the cheapest listed model, priced by the pools: a million tokens in and a million out. - [Fees](https://tokex.dev/docs/fees): Three numbers, each in the quote before you confirm: 5% to buy by card, 0.30% to the pool and 0.05% to Tokex on a swap, and the model’s listed rate to spend. - [API basics](https://tokex.dev/docs/api): JSON over HTTPS under /v1, a tk_ key as a Bearer credential, an Idempotency-Key on every write, and amounts as strings. - [Autopilot](https://tokex.dev/docs/autopilot): Ask for model auto, or auto with a tier, and the cheapest model your credit covers runs, priced through the live pools. - [Caps and panic](https://tokex.dev/docs/limits): A cap is a refusal, not a warning: the request that would cross it never reaches a provider. Panic freezes everything that moves money. - [MCP tools](https://tokex.dev/docs/mcp): A stdio server that is a client of the HTTP API. Writes take an operationKey: the same key returns the first result and never runs twice. - [Endpoints](https://tokex.dev/docs/endpoints): Every route, grouped. JSON over HTTPS with a Bearer credential; the Index answers without one. - [Errors](https://tokex.dev/docs/errors): Every refusal has a status and a stable code. Caps and freezes refuse before any provider sees the request, so a refusal costs nothing. ## Privacy No cookies, no analytics, no third-party scripts or fonts on this site. The browser keeps the Light or Dark choice, the Terminal's chart layout, and, after someone joins the waitlist, their place in line so that browser can show it.