Four layers.
One control plane.
Interface, model access, context and hosting, all in your own cloud account. Any one can change without moving the others: a model change is a setting, a vendor change is not a migration.
What runs where.
Everything Knowbl ships sits inside the rectangle. The only things that cross it are calls to the model endpoints you registered and connections to the systems your people already use, each with that person's own credentials.
The Nexus web app and the Nexus app on computers and phones, under your name. Passwordless sign-in by email code or Google sign-in.
Home, AI Mode, Chats, Messages, Tasks, Dashboards, Projects, Teams, Channels
Every stage points at a tier. The workspace decides what the tier means.
The company's accumulated way of working with AI. The most expensive thing to rebuild, and the thing a fused stack holds hostage.
Your account, your keys, your audit log. Retention, compaction and purge windows are settings.
Set it once. Govern everywhere.
CRM, email, calendar, Drive, ticketing, data. Reached through Tool Connections over MCP. Each user authorizes with their own account; one person's token is never reused for another.
AWS Bedrock, Google Vertex, direct provider APIs, any OpenAI-compatible endpoint including self-hosted vLLM and Ollama, or a custom HTTP endpoint. Credentials encrypted at rest and never returned by the API. Under the standard Nexus architecture, prompts and documents are not used to train public models.
The model is a setting. Try it.
Every stage in Nexus points at a tier, and the workspace decides what the tier means. Remap a tier below and watch what follows, and what does not.
The registry accepts AWS Bedrock, Google Vertex, direct provider APIs, any OpenAI-compatible endpoint including self-hosted vLLM and Ollama, or a custom HTTP endpoint. Credentials are encrypted at rest and never returned by the API.
Nexus, under your name
Registry, tiers, per-stage
Skills, knowledge, memory, connections
Your own cloud account, your agreement
A repricing is a settings change.
Most enterprises buy the interface, the model and their accumulated context as one thing. Here is what happens to each kind of stack when something changes.
| Event | Fused stack (interface, model and memory from one vendor) | Detached stack (Nexus) |
|---|---|---|
| A provider raises prices or moves to usage-based billing | Absorb it, or migrate everything | Route agents to another model in the registry. Tighten token limits. Negotiate with a switch in hand. |
| A cloud credit program ends | Re-platform, or pay list price | Nexus runs in your own cloud account under your own agreement. The compute deal is yours to negotiate; nothing in the workplace depends on ours. |
| A better or cheaper model appears | Wait for your vendor to adopt it | Add it to the registry. Switch it on per agent, starting with routine work. |
| A model lacks a capability | Rewrite the workflow | The router gates by capability and drops what the model cannot do instead of erroring. |
| You want open-weight models on your own GPUs | Rebuild against a new SDK | Register a vLLM or Ollama endpoint as OpenAI-compatible, map it to a tier. Credentials encrypted at rest. |
| A vendor retires a feature | Rebuild, on the vendor's schedule | Skills, Knowledge and Connections stay in Nexus, unchanged. |
| A model provider has an outage | Everything stops | Switch models. The workplace, the context and the connections stay intact. |
The honest version: a repricing becomes a settings change, not "switching is free." Evaluation and tuning are real work; Nexus makes a model change a project in days, not a re-platform in quarters. We cite reported repricing with dates and never predict a named vendor's next move.
What leaves your account. What never does.
- Model calls to the endpoints you registered, through tiers you control
- Tool calls to your own systems, each with the user's own authorization
- Guest experiences you publish, deny-all origins and no tools until you say otherwise
- Skills, knowledge, memories, connections and prompt templates
- Your agents, their schedules, their runs and their owners
- Your audit trail, token accounting, retention and purge windows
- Your data, prompts and outputs, under the data promise
The full data promise, with its qualifier, is on Security and Trust. We do not extend it.