Frequently asked questions

An account on this site and an API key from at least one AI provider, such as OpenAI, Anthropic, Google or OpenRouter. If you want to run your own copy you also need a web host with PHP 8.4 and MariaDB. The installation guide lists the details.

No. It is a workspace, not a model provider. You connect your own accounts and pay those providers directly. Nothing is added on top of their prices.

No. Consumer chat subscriptions do not come with API access. Providers issue API keys from a separate developer console and bill them by usage. Some coding subscriptions, such as ClinePass and Command Code, do offer an API, and those have presets.

No. Upload the release ZIP, extract it, point the site at the public folder and open the address in a browser. The installer checks the server, asks for the database details and creates the administrator account.

There are presets for 25 providers and you can add any service that speaks the OpenAI Chat Completions, Anthropic Messages or Gemini format. The providers page has the full table.

Model lists are fetched from the provider when you connect and when you press Refresh. If the provider has no list endpoint, Max Harness uses a built-in list that can fall behind. In both cases you can add a model ID by hand.

Yes, through the Ollama preset or a custom OpenAI-compatible server. The model server has to be reachable from the web server, and the administrator has to allow local network providers in the admin settings.

Each preset is covered by automated tests with simulated provider responses, and every endpoint was checked to exist. Conversations with real paid credentials were not run for every provider. Use the Test button after adding a key to confirm that yours works.

They are estimates. Token counts come from the provider. The price comes from OpenRouter's model list, Anthropic's published prices or a number you enter yourself. Cached tokens, batch discounts and negotiated rates are not reflected. When no reliable price is known, only token counts are shown.

Yes. A comparison sends the same prompt to each selected model, so you pay each provider for its own answer.

Only your account, through the workspace. The admin panel does not display conversations, files or keys. Someone with direct access to the server or database could technically read stored messages, which is true of any self-hosted software, so choose a host you trust.

The messages in the current conversation, any attached files, your project and assistant instructions, enabled memory items and, when tools are on, tool results. It goes to the provider of the model you selected and nowhere else. The privacy guide goes through each item.

Encrypted in the database with the installation's application key. They are decrypted only on the server at the moment a request is made. After saving, the workspace shows the last four characters and nothing more.

Open Settings, then Data and privacy. You can export everything, delete all conversations, remove provider keys or delete the whole account. Deletion is immediate and cannot be undone.

Search the web (if a search service is configured), read public web pages, save a note to your memory and create a text file for download. It cannot run code, access the server or see other users' data.

Only for tools you set to Allow, and only search and page reading can be set that way. Saving to memory and creating files always wait for your approval.

Web search needs an external search service. Add a Brave Search or Tavily key, or the address of a SearXNG instance, under Settings, then Tools. An administrator can also configure one for the whole site.