Concepts¶
How an agent built with a Fadenstack SDK divides the work between the device and the server.
An agent runs inside other software: a desktop application, an Office add-in, a browser extension, a web page. The SDK runs the agent's loop on the device. The Fadenstack server governs it.
| On the device (the SDK) | On the server (Fadenstack) |
|---|---|
| The conversation and its history | The agent's instructions, model and knowledge |
| The host's tools, which act on the open document, page or application | Tools on the server: MCP servers registered by IT |
| Asking the user before a tool changes something | Who may use the agent, and on how many devices |
| The tool mode the user picked for the session | The agent's contract: which host tools it may offer, and how strictly |
| Personal-data redaction, limits, budgets on servers that set them, usage and audit |
The rest of this section describes each part:
- Agents on the server: how an agent is registered, versioned and granted to people.
- Tools and approvals: tool classes, tool modes, and where tools run.
- Outcomes: how a turn ends, including when something is refused.
- Sessions: where conversations are kept, and what the server sees.
- Personal data: what the model sees when the organisation redacts personal data.