Product and dev teams
Parallel coding sessions, reviews, deployments and operations, all supervised. The kanban becomes a queue of sessions.
SOKKAN is the open-source agentic studio where AI agents move your work forward, where nothing important ships without your click, and where everything they learn joins one memory, recalled on every exchange and reviewing itself. On your own machine, or operated for you in Switzerland.
Who it is for
Born to run our own products, SOKKAN serves any team that wants to hand work to AI agents without losing control, or the memory of what was decided.
Parallel coding sessions, reviews, deployments and operations, all supervised. The kanban becomes a queue of sessions.
Files, templates, procedures: every decision and convention stays in memory, dated, and comes back when an agent needs it.
Open source and self-hosted, or operated in Switzerland. Self-hosted, the memory runs on your machine and never leaves it.
One cockpit to move projects and operations forward without hiring an ops person: Nina, the onboard assistant, answers in the interface.
New in SOKKAN 3.0
In 2.x, memory arrived when a session started. In 3.0 it arrives with every message, in every session and every sub-agent, with the age of each note and where that date comes from. And it reviews itself.
A session opens from a kanban card with the project’s relevant notes already in its context. Starting a task already means “check your memory”.
With every message, SOKKAN looks up what the question touches and adds it, with the date of each note and where that date comes from. An agent no longer knows at the start and forgets along the way.
When an agent delegates, the sub-agent gets the recall for its topic as well. Every recall is logged: who received which note, and when.
The CortHeXis tab scores memory health from 0 to 100: broken links, duplicates, descriptions that no longer match the content, forgotten work. Every fix is one click, and waits for your approval.
Measured, not promised
82%
of questions get the right note as the first result with SOKKAN 3.0 on a GPU, against 42% with SOKKAN 2.x. Without a GPU, on a small server: 73%.
Benchmark of 300 questions in French, English and German, asked of our own working memory (415 notes), 2–3 October 2026. “First result”: the expected note comes out on top; MRR: mean reciprocal rank (1 = always on top). Method and details.
Magnitude picks
At install time, Magnitude looks at the machine and picks the memory profile. Switch whenever you like: the new index builds in the background while the old one keeps answering, and the switch only happens if the measurement on your own notes does not regress. The old index is kept for 7 days so you can roll back.
4 cores · 4 GB · no GPU
73%first result
8 cores · 16 GB · no GPU
73%live, better deferred
A GPU, or the ANCHOR appliance
82%first result
The recommended memory model, Google’s EmbeddingGemma, is free, commercial use included, but it is not open source. So SOKKAN does not ship it: it downloads it on first launch, once you have read and accepted Google’s terms. It runs entirely on your side; no data is sent to Google. Prefer not to? SOKKAN uses an MIT-licensed model instead, with no loss of features. Full details.
The CortHeXis tab
Memory fails silently: a link to a renamed note, two notes saying the same thing, a date that lies. In SOKKAN 3.0, the CortHeXis tab reviews the memory, scores it and proposes the fix.
The studio
Agent tools often stop at delivery. SOKKAN keeps going: it operates production with you, and the green light stays human, wherever you are.
Agents’ permission requests and questions become buttons. Nothing irreversible without your click, nothing marked done without you.
A production alert opens a session that knows the project, proposes a fix and waits for your green light, with a Telegram ping to come and give it.
Metrics and logs inside your network; your sessions read them and compose a dashboard on request.
Encrypted and injected into sessions as environment variables: the agent uses them without the value ever showing, on screen or to the model.
A procedure is a memory note. “Replay” opens a guided session, step by step, with human sign-off on anything irreversible.
Deploy an image to a worker in your fleet, roll back to the previous version in one click. Per-session costs, an audit log of every action.
Nina
Running a platform should not take a full-time ops person. Nina knows SOKKAN inside out (sessions, memory, fleet, runbooks) and answers in the cockpit, right next to your work. Typical conversations:
api-5xx alert has already opened a diagnosis session, loaded with the project memory: the payments worker has been running out of memory since the 10:14 pm deploy, and it proposes a rollback. Nothing ships without your go.decision-sqlite-storage note, in May. Mark it as a priority and every session will see it before proposing anything.scraping-refactor session got stuck in a retry loop. Move that task to a small model and replay the “rate limits” runbook.Included in every SOKKAN Cloud instance (the 🧭 button in the cockpit); as soon as she has access to your context, her inference must run in Switzerland, and the content of your exchanges is never kept or used to train anything. Self-hosted, enable SOKKAN_FEATURE_ASSISTANT with the model of your choice. A question about SOKKAN? Nina answers here too, bottom right.
Free, no limits
Docker, plus your Anthropic key or your Claude Pro/Max subscription. Everything stays on your machine.
$curl -fsSL https://sokkan.ch/install.sh | sh
Requirements: Linux (or macOS with Docker Desktop), about 4 GB of RAM, an Anthropic key or a Claude Pro/Max subscription. No Docker? The script offers to install it (official get.docker.com script).
SOKKAN Cloud
The same SOKKAN, minus the ops: a dedicated VM in the Geneva zone, your own private network, your URL at you.sokkan.ch. Bring your Anthropic key or take managed inference, served from our own GPUs in Geneva (fallback in France, or a 100% Swiss tier), prepaid credits from CHF 25 (+25% credits from CHF 500), metered per token. Free 14-day trial on every plan: card required, first charge on day 14, cancel in one click before (trial seats are limited).
Grow from the cockpit: worker instances from CHF 59/mo · managed PostgreSQL from CHF 99/mo (daily backups), inside your private network. Web exposure included: HTTPS subdomains and your own domains, automatic certificates, at no charge. Prorated billing, provisioning after payment, cancel monthly.
Hosted at Exoscale (Geneva), operated by Ninabot Sàrl: your data stays in Switzerland. Not sure about sizing? [email protected], you’ll get the founder, not a bot.
In the press: ICTjournal, “Ninabot launches a Geneva-hosted cloud for AI agents” (in French, 04.08.2026).
Go further