Keeping client strategy data in your own environment
If your business is built on confidential method — a consulting practice, a law firm, a family office, an M&A boutique — then "where does the data live?" is not a procurement checkbox. It is a question you answer to your own clients, and a wrong answer costs you the engagement.
Most strategy platforms answer it with certifications and hosting regions. Those are real answers to a real question, and for many buyers they are sufficient. They are not responsive to the objection some buyers actually raise, which is architectural: no permanent copy of our clients' strategy on a third party's servers, at any price, under any certification. hoshin.app answers that one directly — the same 14 tools and the same completeness score run against plain JSON files on a machine you control, with no account and no network.
The objection, stated properly
We built the sovereign mode before we met the objection, and then a real prospect articulated it better than we had. A governance and compliance partner at a consulting firm declined a hosted strategic tool on principle of architecture — not as a negotiating position, and not after reading a DPA. His reasoning: the firm's assets are immaterial. Client knowledge, methods, the shape of the questions they ask. That is trade-secret material, and its permanent residence at a vendor is a category of risk he does not accept for a management tool.
He framed it as a segment prerequisite, not a personal preference. Any practice living off confidentiality has the same reflex. Encryption at rest does not address it, because the objection is not about interception; it is about custody.
Four levels, one floor
Residency is not a checkbox, it is a ladder — and a buyer should be told which rung they are standing on.
The floor, whatever the level: no use of your data to train models. That one is structural rather than promised — the language model in the loop is *yours*, reached through your own MCP client. We do not call a model on your behalf, so there is no inference path through which your strategy could be used for training.
- Level 1 — hosted: your organization on hoshin.app. EU hosting, strict per-organization isolation on every query, an audit trail on privileged operations. The classic SaaS relationship.
- Level 2 — hosted, published selectively: the same, plus revocable read-only links for people outside your organization. You choose what leaves and you can withdraw it.
- Level 3 — your own model: the assistant is yours. Your Claude, your Cursor, your API key. We never sit between you and a model provider.
- Level 4 — sovereign local: plain JSON files in a directory you chose, on your machine. No account, no network, no telemetry. The MIT-licensed code is auditable before you run it.
What level 4 actually guarantees
It is worth being precise, because "local" is used loosely in this market.
There is no HTTP client for your data in the local transport. Nothing is sent because there is nothing to send. There is no account to create, nothing phoning home, and no analytics. The files *are* the export — one JSON file per matrix, readable, diffable, and versionable in your own git repository, which also means an Article 20 portability request is satisfied by a copy command. Deletion is `rm`, not a support ticket and not a thirty-day retention window.
The engineering that matters if you are trusting it with the only copy: writes are atomic (temporary file, then rename), serialised through an in-process mutex so concurrent tool calls cannot interleave, and split one file per matrix so a bad write cannot take the whole dataset with it.
And the responsibility that comes with it: backups are yours. Local mode moves custody to you; it does not remove the duty of care.
The one capability sovereignty costs you
Publishing a public share link is refused in local mode, and the tool returns an error saying so. This is not an unimplemented feature. Publishing needs a server to publish to, and quietly standing one up would undo the guarantee the mode exists to provide.
The substitute is better suited to the segment anyway. Ask your agent for a matrix artifact: it renders a self-contained interactive snapshot of the strategy that you send deliberately, once, to a recipient you chose — rather than a URL that keeps serving your client's strategy until someone remembers to revoke it. For a firm, a one-shot artifact is easier to defend to a client than a live link.
Why the business logic is identical in both modes
A sovereign mode that lags behind the hosted one is a marketing claim, not a feature. So the scoring engine is pure — elements and correlations in, a verdict out — and a test compares the hosted implementation against the local one on the same matrix, asserting an identical score, an identical level and an identical list of suggestions. If the cloud version ever gains a rule the local one lacks, the build fails.
The same rule governs the tool contract: 14 tools, one shared definition, and a test that every transport implements every one of them. Your client's strategy is judged the same way whether it sits in our database or on your laptop — which is the only way a firm can defend using the sovereign mode without telling clients they got the lesser product.
What we do not claim
We are a small product and we would rather you learn our limits here than in a security review. We do not hold SOC 2 Type II or ISO 27001. We do not publish a signed DPA template on this page today — ask us and we will discuss the specifics of your engagement. Our sub-processors are limited: hosting and transactional email. Nothing else touches your content.
If your client requires certifications from every vendor in the chain, a larger platform will clear that bar and we will not. If your client requires that no vendor holds the data at all, we are one of the only options that can say yes.
Frequently asked questions
Can I use hoshin.app without creating an account at all?
Yes. The sovereign local mode is the default binary of the npm package: `npx -y hoshin-kanri-mcp` with a data directory. No account, no network, no telemetry. Your assistant connects to it over stdio on your own machine.
Can I run one client per environment?
That is the normal pattern for a firm in local mode: one data directory per client engagement, kept wherever that client's material is allowed to live. Nothing is shared between directories because nothing is centralised.
Do you use our strategy to train models?
No, and it is structural rather than a promise: the model is yours, reached through your own MCP client. We do not call a model on your behalf, so there is no path by which your content could reach a training set through us.
Is the local mode a free tier in disguise?
It is not a degraded tier. Same 14 tools, same completeness score, same guided prompts, enforced by a parity test. The single exception is public share links, refused by design because publishing requires a server.
What happens to my data if you shut down?
In local mode, nothing — the files are on your machine and the MIT-licensed code is in a public repository you can keep running indefinitely. That is the honest answer to vendor risk for a small product, and it is why the code is open.