How SOKKAN works
Three diagrams, generated from the code and checked against it. Click a box to see what it does and what it talks to; press Live in the full view to trace the main path; Export gives you PNG, SVG or a WebM clip. Version française.
1.A session, end to end
You open a card or type a prompt. The API runs the memory search itself and writes the top notes into the session's first message, so the agent starts already knowing what previous sessions learned. The session runs on the Claude Agent SDK; every tool call that would change something goes through the HITL gate and comes back to you as a button. All of it lives in one container on your machine.
Every box marked SRC links to the file and line in ninabot-ch/sokkan at the commit the diagram was generated from. If the diagram and the code disagree, the code wins and we regenerate.
2.From a card to an approved change
The loop you will run every day. The part that matters is the order: recall before any code, a plan grounded in notes and code, then a gate on every mutating step. Denied means the agent replans; it never bypasses. A card only reaches Done when a human moves it.
3.An agent run (Crew)
Crew agents run on their own: a cron, or a production alert. A run is still an ordinary session with the project memory, the agent's allowed tools, vault secrets by name and a per-run budget. It pauses the moment it wants to write, stops at 100 % of its budget, opens an incident if it fails, and the notes it writes stay quarantined until you promote them.
4.Going further
- First steps — from login to a working memory in 20 minutes.
- Crew agents spec, Operate, multi-user — on GitHub.
- Diagrams made with archify (MIT). Sources are in the repository and regenerated at each release.