How the agent surface is built

stella’s design principle is that everything a human can do in the app is equally available to scripts and AI agents, under the same permission model. Instead of maintaining three integrations, stella maintains one tool registry and projects it three ways:

• the remote MCP server advertises the registry to connected assistants;
• the CLI is code-generated from a committed registry snapshot, so stella <resource> <action> mirrors the tool surface;
• the in-app assistant works from the same registry through a sandboxed execution layer.

Why it cannot drift

Each projection is guarded in CI: the committed registry snapshot, the capability catalog, and the generated CLI tree are all diffed against the live registry on every change. A tool added or changed in one place is a reviewable diff in the others, never a silent divergence.

Permissions

Access is scoped twice: OAuth scopes decide which tools a connection can see and call, and every call is authorized against the organization and workspace it touches, the same checks the app itself uses. Destructive tools additionally require an explicit confirmation from the caller.

Self-hosting

A self-hosted stella serves the identical surface: the endpoints and CLI in these docs work against your own instance by swapping the host.