Getting started

From nothing to a first answer from a model. The same account works for everything else on the platform: renting a machine, renting yours out, or putting one into a model pool.

Get an invitation

During the test, accounts are created from an invitation, not from a sign-up form. Invitations come from our Telegram bot.

1. Message the bot

   Open the bot and ask for an invitation link. Find the bot here.

2. Say what you came for

   Use the network, host a machine, or both. It only tailors what the bot tells you next — every account gets the same rights.

3. Get the link

   The bot replies with a one-time link. It works once and expires within hours, so open it soon after it arrives; ask again if it lapses.

Some deployments open self-registration instead. If the sign-in page shows a registration tab, you can use it; ours is closed during the test.

Create the account

• The invitation page asks for a username and a password of at least 10 characters. Nothing else is required — no email, no phone, no payment details.
• You accept the Test Terms and the Acceptable Use Policy. If you later rent your own machine out, the Host Terms are added.
• When a document changes, the panel asks you to accept the new version before you continue. There is no way to withdraw an acceptance yourself — ask us in Telegram.

After that you sign in at the panel, which is where everything else happens.

Test credits

Everything is settled in test credits: model requests, machine rentals and what a host earns. Credits have no monetary value. Nothing is charged to you and nothing is paid out.

• How you get them. The team grants them to your account. There is no way to buy or top up credits Planned.
• What spends them. Tokens you send and receive through the gateway, and every minute of a rented machine.
• Where you see them. The Developer console in the panel: the Credits tab shows your balance, what is held in reserve and your last entries; the Usage tab breaks requests, tokens and credits down by day and by model over 7, 30 or 90 days.
• When they run out. Gateway requests stop being served, and a rented machine is stopped on its host rather than quietly kept running.

Your first request

The gateway speaks the OpenAI API, so the official SDKs work unchanged — you point them at our base URL and use a ServerShare key. In testing

1. Create a key

   Developer console → API keys. The key is shown once, when it is created or rotated; we keep only its hash, so a lost key is replaced rather than recovered.

2. Pick a model

   GET /v1/models lists what the pools serve right now, with the context length and the capabilities of each model.

3. Send a request

   POST /v1/chat/completions, with stream: true if you want tokens as they are produced.

curl https://servershare.io/v1/chat/completions \
  -H "Authorization: Bearer $SERVERSHARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "example-chat-model", "messages": [{"role": "user", "content": "Hello!"}]}'

example-chat-model is a placeholder — take a real name from GET /v1/models. The full reference, including errors and the fields we add, is in the API docs.

Chat completions, image generation, the model list and asynchronous jobs are the /v1 surface today In testing. Embeddings and speech are Planned.

Chat Studio In testing

If you would rather not write code, the panel has a chat. It is a client of the same gateway, so it counts against the same credits and the same limits.

• It issues itself a key that can only list models and chat, keeps it in the browser tab's memory and replaces it within the day. You can see and revoke it in the Developer console.
• Conversations are stored on the server, so they follow you between devices; you can rename, export and delete them. Deleting a conversation deletes its replies too.
• You can put two models side by side. That is two requests, so it costs twice as much.
• Attachments, search across conversations and shared links are Planned.

Limits

• Each API key can carry its own requests-per-minute and tokens-per-minute limits, set when you create it. A key without them is bounded only by the account limits.
• Your account has its own requests-per-minute and tokens-per-minute ceilings. They are the same for everyone during the test; per-account limits you can change yourself are Planned.
• A short burst well above your usual rate is refused with a 429 rather than queued.
• You may hold up to 20 active keys at once.

Your data

• Export. The panel exports everything the platform holds about your account as one file. Secrets are not in it: keys appear by name and prefix, never by value, because the value is not stored anywhere.
• Deletion. The panel first shows what would be erased and what would be kept, then asks you to type your username to confirm. It cannot be undone.
• What is kept. Records that belong to somebody else's accounting as much as to yours — credit entries, rental contracts, the request log — stay, with your account reduced to a stub that no longer identifies you.
• When it is refused. While a rental is running, a contract is awaiting your decision, a machine is still up or in a pool, or your balance is negative. Close those first.

Requests may be served by third-party model providers; which one served a request is shown next to it in the Usage tab, and the arrangement is described in the Test Terms.

Where to go next

• Rent a machine — a full virtual machine with a GPU share, for work the gateway does not cover.
• Become a host — put your own machine to work.
• Model pools — serve a model to the whole network.
• API reference — every endpoint, field and error.