LakkaA self-hosted AI control plane for teams of 10 to 50 people

Use its chat, or point the apps your team already uses at its OpenAI-compatible API. Either way, spend cannot exceed a cap, every request is attributed and traced, and a tool can never do what the person asking is not allowed to do.

One set of rules for every surface
Web chat, the OpenAI-compatible API and MCP tools check one set of people, permissions and budgets. A key acts for the person who made it.
Workspaces and people
Caps that hold
Every cap in the chain is reserved before a request is sent. If the stores cannot be reached, the request is refused.
Budgets and spend
Routing to the right model
Ask for the model auto, from chat or from any app, and the workspace's routes pick the model for each request. The classifier that matches them runs offline inside the image.
Routing
Traced, and tools that keep to the rules
Every request is traced with its cost and routing decision. Tools run as the person who asked; command-line servers run in a sandbox.
Tools and MCP servers