Building with external assistants

Connecting an assistant

1. Open your profileIn Profile choose Connect an assistant.
2. Choose organization and nameSelect the organization and give the client a name, for example "Claude Desktop on my laptop".
3. Create the tokenEach assistant has its own token, revocable at any time. Its creation is recorded.

How it works

1. Understand what is thereIt reads the organization (secret and connection names, models, sites, services), the recipes and the existing projects.
2. Read before writingFor each node type it reads the guide, fields and ports, without guessing names.
3. Start from a recipeWhen a suitable one exists.
4. Build in coherent piecesOne channel, its agent and its tools at a time, with the option of a dry run without applying.
5. VerifyAfter every change it checks the project and fixes what it reports.
6. Go back if neededIf a build fails, status and logs are analyzed; rollback is the way out.

For code, the assistant first reads the workspace guide, edits with exact replacements, checks the types and records a commit for every coherent change.

Safeguards

• No assistant tool deletes a project.
• Removing a node with data, changing its type or rolling back beyond the point where it was added is rejected without explicit confirmation; the refusal says what would be lost.
• Secrets are never requested, written or repeated: the node references them by name and the assistant says which name to set in the console.
• Public-facing agents are not given full access, deletion or code editing.