Radar / Ideas / Sovereign Desk: a private knowledge…

Sovereign Desk: a private knowledge agent that runs on one workstation

daily ideaambitiousJEV confidence 0.422026-10-10
OutcomeA 50-person firm gets a frontier-quality private knowledge agent running on a single workstation, with no document ever leaving the building.

The problem

Law firms, clinics, and industrial companies sit on documents they cannot send to cloud AI under client or regulatory policies. Current local stacks (Ollama plus small open models) keep data in-house but answer like a weak model, so teams quietly keep pasting into ChatGPT anyway. Demand is real but the tooling is DIY: TeleAgent just shipped a private-deployment edition for sovereign environments, and GitHub hosts multiple sovereign-AI reference projects trying to solve exactly this.

The idea

Ship a preconfigured Docker Compose bundle pairing WeKnora's self-hosted knowledge platform (RAG, agent, self-maintaining wiki, per-workspace RBAC) with Colibri's memory-multitiering inference engine, so a 744B-class open MoE runs on one workstation with a fast NVMe drive. A decision router sends trivial lookups to a small local model and reserves the big MoE for hard questions, keeping interactive latency acceptable. The MSP installs the box, loads the firm's documents, and hands over a working private agent.

Why now

Two unlocks landed within days of each other. Colibri demonstrated 744B-2.8T-class MoE models running on consumer hardware by treating VRAM, RAM, and storage as one inference hierarchy. WeKnora shipped a production-grade self-hosted knowledge stack (29 model vendors, swappable local backends, Langfuse tracing) that already supports local Ollama. Together they turn a frontier-quality private knowledge agent from a GPU-cluster project into a workstation purchase.

What it combines

Colibri's AI memory multitiering (giant models on consumer hardware) plus WeKnora's self-hosted knowledge platform (RAG plus agent plus self-maintaining wiki). The mix matters because WeKnora solves the knowledge-management half that Colibri lacks, and Colibri solves the model-quality half that WeKnora's local-Ollama path lacks; neither alone delivers a frontier-quality private agent on cheap hardware.

MVP

Docker Compose bundle: WeKnora plus Colibri serving one supported MoE (e.g. GLM-5.2), a setup wizard that ingests a folder of PDFs, and a published benchmark of answer quality and 5-concurrent-user latency on a mid-range workstation. Deliberately skip multi-tenancy and fine-tuning.

Distribution

B2B2C through MSPs and IT services firms that already sell into law firms, clinics, and industrial companies. They charge for the workstation, the setup, and a maintenance contract; the software bundle is the wedge that wins the engagement.

Why it wins

AnythingLLM and Open WebUI plus Ollama already serve small firms, but they cap out around 70B-class models on a workstation. GrapeUp Aiboostr and SecureLLMs sell private AI as managed per-seat services. Sovereign Desk is neither: a one-time hardware-plus-software package, sold through MSPs, that brings 744B-class brains to the same box the firm already owns.

Risks

Interactive latency: Colibri's honestly reported numbers are slow on weak hardware (fractions of a token per second in streaming CPU mode). The MVP de-risks this by publishing measured latency on a defined hardware spec before selling anything; if it cannot beat 5 tok/s on a $2,000 box, the product thesis fails.

Build it with

Repo to start from

sovereign-desk: Docker Compose bundle plus setup wizard and published benchmarks for private knowledge agents on consumer hardware.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.