Every company will run its own AI platform.

A general chatbot answers trivia. A private platform answers your questions — “what’s the status of order 4471”, “draft the renewal email for this client”, “summarise last week’s support tickets” — because it can reach the systems that hold those answers. We wire the assistant in the way we do it every day, using the Model Context Protocol and clockwork, our own open-source Go MCP server.

Most private-AI setups stop at a chat box: your team asks a question, the assistant types a paragraph back. We go one step further. When the assistant looks up your orders, invoices, or customers, it can hand them back as an actual interface — a sortable table, a form you can edit and submit, a grid of cards — rendered inside the conversation. Fewer copy-paste steps, fewer mistakes, real work finished in the chat.

This is not a bespoke trick that breaks on the next update. In January 2026, Anthropic and OpenAI standardised exactly this as MCP Apps (SEP-1865) — the first official extension to the Model Context Protocol. It lets an MCP server ship interactive UI that renders inside any compliant host: Claude, ChatGPT, Cursor, and the private workspace we build for you. Building on the open standard means the interfaces your assistant shows are sandboxed, portable, and tied to no single vendor.

The same way clockwork gives the assistant secure tools over your systems, our new open-source library gomukit gives those tools a face. It generates the MCP UI widgets — data tables, card grids, and validated forms — as self-contained HTML that themes itself to the host and ships without external dependencies. It is the engine we use to put a real interface over your data.

This is the part most “private ChatGPT” setups skip. Even when the platform is yours, every prompt your staff type can carry customer names, credentials, or identifiers straight to a model provider. We put a privacy gateway in front of the models — using Microsoft Presidio — that detects and redacts sensitive data before the request leaves your boundary.

The redaction policy is configured per deployment: which entity types to catch (personal data, credentials, account numbers, whatever your domain requires), what to mask, and what to allow. It is inspectable — you can see exactly what the gateway is doing. Privacy here is an architecture, not a promise on a pricing page.

Most companies want the best models and a strong privacy posture. Some cannot let data leave their infrastructure at all. We serve both, and you can start on Standard and move workloads to Sovereign later.

This is a managed service, not a one-off install we hand over and disappear from.

The same per-milestone, stop-anywhere model as every Techthos service.