Chat interfaces¶
People will chat with AI whether or not you give them a tool for it. Give them one that runs through the gateway and every message is checked against your policies, charged to the right team and recorded in Request logs, on GPT, Claude and Gemini models alike. Each app below needs two values: the gateway's address and a Vidai key.
Pick your app¶
| App | Who runs it | The settings to change | Each person's own key | Page |
|---|---|---|---|---|
| Open WebUI | Your server, for a team | Admin Panel › Settings › Connections › OpenAI API: URL, API Key | Yes, with Direct Connections | Open WebUI |
| LibreChat | Your server, for a team | A custom endpoint in librechat.yaml: baseURL, apiKey |
Yes, apiKey: "user_provided" |
LibreChat |
| AnythingLLM | Desktop, or your server for a team | Settings › AI Providers › LLM, Generic OpenAI: Base URL, API Key | Desktop only | AnythingLLM |
| Chatbox | Each person's device | A custom provider: API Host, API Key | Always | Chatbox |
| Jan | Each person's computer | A custom provider: Base URL, API key | Always | Jan |
| Msty | Each person's computer or browser | An OpenAI-compatible provider: endpoint, API key | Always | Msty |
The gateway's address is under Connect your app, on the Keys page and in System › Settings › Gateway. Keys come from Keys › Issue a key; a key's secret is shown once.
For every chat app¶
- Shared deployments use one key. When an admin sets the key on a team server, every person's messages go out on that key: they show under it in Request logs and are charged to the team it acts for. Where the app lets each person enter their own key, they issue one in Keys › Issue a key and their calls show under them, charged to their own team. Otherwise, to split cost by team, give each team its own connection or instance with its own key.
- The model list is the key's. These apps fill their model picker from
GET /v1/models, which the gateway answers with exactly the models the key may call. Approve a model under Models in the Control room and it appears after the app refreshes its list. - Use the OpenAI chat format. Chat is translated to every provider. An app's image, speech or embedding features reach OpenAI and Azure OpenAI models only, and an app's "Responses" mode reaches OpenAI models only.
- Background calls count. Titles, tags, follow-up suggestions and agent steps are calls too, each with its own row and cost in Request logs. Point them at a small model where the app lets you.
- Find a call. Chat apps don't show the gateway's response headers. Look the call up in Request logs (filter Key, Model or Path) within seconds of being made.
- Guardrails apply. A message a guardrail stops comes back as an error in the chat, with the guardrail's message. See Headers and errors.