Introduction

Kai is an AI chat app that you run yourself, for any model. A team signs in, creates a workspace, adds its own keys (for a provider directly or for a gateway), and chats with the models and with the tools it connects over MCP. Conversations, keys and files stay in your Postgres and your object storage.

What you getCopy Anchor Link

• Workspaces and roles. Owners, admins and members, with invitations. Each workspace has its own keys, tools, agents and usage.
• Bring your own keys. Anthropic and OpenAI, or any compatible gateway (Azure OpenAI, LiteLLM, an internal proxy). Kai knows each model's context window, reasoning efforts and features, and switches them on for you.
• MCP tools. Remote MCP servers (Streamable HTTP or SSE), local commands (stdio), and plain HTTP APIs described by an OpenAPI document. Header, OAuth and client-credential authentication; values shared by the workspace or filled in by each person.
• Approvals. Kai asks before a tool runs, unless the workspace allows it. People can allow a tool once, for the chat, or always.
• Agents. Instructions with variables like the person's name and today's date, plus the tools and models a chat may use.
• A careful chat. Streaming that survives a page reload, editing and regenerating as branches, attachments (images, PDFs, text), web search, reasoning effort, and automatic summarizing when a conversation gets long.
• Built for the prompt cache. Each request starts with the previous one, byte for byte, so long chats and tool loops are mostly read from cache.

Where to startCopy Anchor Link