One admin, many models, no enterprise bloat

What it does

HiveChat is a self-hosted AI chat interface built for small and medium teams. An administrator connects API keys for providers like DeepSeek, OpenAI, Claude, Gemini, and a long list of domestic Chinese services, then decides which user groups can access which models and how many tokens they burn per month. Everyone else just logs in and chats.

The interesting bit

It treats LLM access as a shared utility rather than a collection of individual API keys. The integration with Chinese enterprise auth—WeCom, DingTalk, and Feishu—and support for MCP servers in SSE mode suggest the authors actually run this inside real organizations, not just side projects.

Key highlights

• Supports 15+ model providers out of the box, including OpenAI, Claude, DeepSeek, Moonshot, Qwen, Doubao, and Ollama, plus any custom OpenAI-compatible endpoint.
• Group-based access control with per-group model visibility and monthly token quotas.
• Enterprise login options: email, Feishu, WeCom, and DingTalk.
• DeepSeek reasoning chain visualization, image understanding, LaTeX/Markdown rendering, and AI agents.
• Built on Next.js, Auth.js, PostgreSQL, and Drizzle ORM.

Caveats

• Docker deployments currently lack automatic database migration scripts; upgrading between versions often requires manually running SQL ALTER statements.
• The README warns that rapid recent updates have made Docker upgrades fragile enough that the author suggests wiping the database volume for test environments or contacting them directly for production help.

Verdict

Worth a look if you lead a small company, school, or department that needs gated LLM access without paying enterprise SaaS rents. Skip it if you need fully managed, zero-touch upgrades or deep analytics beyond token counting.