Security overview, for your review.
Kelynto Managed runs on Microsoft Azure in Central US, and the same software can run inside your own Azure account. A reviewer needs three things from a vendor, and this page gives them: the short version, every known limit stated precisely, and the way to get the full review pack.
The short version
- Sign-in is yours. OpenID Connect through your identity provider. No passwords are stored. Four roles, checked on every API route. A sign-in by emailed link and one-time code is also built, for workspaces without single sign-on; it is off unless a deployment switches it on, and a workspace can require single sign-on instead. A workspace admin can set up single sign-on and SCIM provisioning in the product.
- Tenants are kept apart twice, in the application and with PostgreSQL row-level security that fails closed.
- A partner's clients are kept apart. Each client of a partner has a workspace of its own. A partner key acts for one client per call, and a client that is not the partner's answers not found. The partner's people sign in through the partner's own identity provider and reach its clients only, and a client that leaves a partner ends the partner's access at once. Every partner request is in the client's audit log.
- Access can be limited by region, district or store. A member who is not an admin, an API key or a webhook can be limited to some stores. The limit is enforced three times: a restrictive PostgreSQL policy on every table keyed by store, the same filter on every query, and a closed list of the routes a limited caller may reach. Every change of a limit is in the audit log.
- Recurring data deliveries can be governed. Control totals reconciled to the cent, source profiles that hold a changed file, held files decided with a reason and not approved by their sender while someone else can, and a delivery history kept as long as the audit trail.
- Every change made through the application is audited.
- The AI writer is a choice per workspace. The engine computes every number. The commentary is written by Kelynto's built-in writer, and the in-app assistant answers from the workspace's records by fixed rules. The AI writer is switched on for a workspace by its admin, or for a partner's clients under the partner's recorded agreement with them. It then receives the week's evidence: store, region and category names, the week's amounts and shares, the titles and dates of records, and the results of the calculations it asked for, with no row, no person, no file and no email address. It is Claude, by Anthropic, unless the admin chooses the company's own model, whose key is sealed and shown only by its last four characters. If the assistant's model mode is switched on, its model receives only the question and the name of the page, and picks the kind of question. A model never writes what a person reads in the assistant.
- It can run inside your own Azure subscription, with the deployment kit that runs Kelynto Managed. Two deployment settings can forbid every outbound signal lookup and every language model call for every workspace.
- Service status is public. status.kelynto.com checks the Kelynto Managed environment from outside every 5 minutes and shows the current state and uptime of the web app, the API and the database, the AI writer's check after each deployment, and any incident.
Each of these, with its status label, is in the Trust center. Where it runs and what crosses the boundary is on the Enterprise deployment page.
What data does it handle?
Weekly totals by store and category. Kelynto does not need or store customer, shopper, payment or employee HR data.
| Data | Sensitivity |
|---|---|
| Weekly sales and plan by store and category | Confidential business data. Totals only: no transactions, baskets or SKUs required. |
| In-stock rates and promotions | Confidential business data. |
| Store list with coordinates | Low. |
| User email, display name and role | Personal data (business contact), supplied by your identity provider, or typed at sign-in where email sign-in is on. |
| Commentary, cause labels and weekly pulse answers | Confidential. |
| Causes planners add, follow-up actions and the history of decisions | Confidential. A file attached to an added cause is type- and size-checked and served only as a download. |
| Questions asked of the in-app assistant | Confidential. Kept per person, readable only by the person who asked, and covered by retention, the full export and the deletion of the workspace. |
| Files received and the delivery history | Confidential. A held file is kept compressed until it is decided, and no longer than the workspace's retention of files; the delivery history is kept as long as the audit trail. |
| Credentials an admin enters for scheduled collection, and webhook secrets | Secret. Stored encrypted, never shown again and never logged. |
| A company's own model key for the AI writer | Secret. Sealed with the deployment's key and shown only by its last four characters. |
Store notes, a few words and an optional file that a store or district manager leaves for one of their own stores, are read and written only for that person's stores, enforced by the database's row policies as well as the queries and the routes. Every write is in the audit log, they carry no amount, and an admin can switch the feature off for the workspace. The Monday digest is a plain email sent through the mail path every other message uses; where no mail relay is configured nothing is sent.
Known limits, stated precisely
For the reviewer who asks: each item is a limit of the software or its deployment templates as they stand today, stated so that nothing surprises you in review. Every control and its evidence is in the Trust center.
- Not SOC 2 certified, and no third-party penetration test yet. The controls are documented one by one in the Trust center, each with its evidence, and an internal security assessment of the release covers cross-workspace access, role escalation, token attacks, upload abuse, injection and server-side request forgery. A customer's own test is agreed in the contract.
- Sign-in, single sign-on set up by a workspace admin, domain verification and SCIM provisioning are tested against stand-in identity providers. The first connection to your provider is checked with a test sign-in and a test group.
- Sign-in routing shows whether an email domain is registered. It never shows whether an account exists.
- Accounts are matched by email. The email claim must come from an attribute your users cannot edit themselves. The token's subject is stored but not checked afterwards.
- Tokens from your identity provider are not checked for revocation. Each token is verified on every request, but one already issued stays usable until it expires. Keep access token lifetimes to an hour or less.
- With single sign-on, refresh tokens are held in the browser tab's session storage, like the access token. Rotation and reuse detection are settings of your identity provider.
- Email sign-in is as strong as the mailbox. Whoever can read a person's mail can sign in as them. The second factor is optional for an account until a workspace requires it, and it is an authenticator code: there are no passkeys and no hardware keys.
- The email-verified claim is not checked for workspace users. On a shared provider that lets people sign up with an address they do not own, restrict who can receive the tenant claim.
- The owner database role is not limited. The application connects as a database role that owns nothing, so it cannot switch row-level security off or alter a policy. The separate owner role, used for migrations, can, and nothing limits it beyond who holds its credential.
- The tenant binding and the store limit are session settings. Row-level security guards against a forgotten filter. It does not contain an attacker who can already run arbitrary SQL as the application.
- Access by store has no limit by category, and the commentary a limited member reads is written for their stores each time and is not versioned.
- The list of tenants is not under row-level security. It holds names and settings, no sales data.
- Record ids are sequential across tenants. A tenant cannot read another tenant's records, but it could estimate overall volume from its own ids.
- The audit log is tamper-evident, not tamper-proof. The application's database role can only add events, and each event is linked into a hash chain per workspace. The database role that owns the tables could rewrite events together with their hashes, which shows only against a copy exported earlier.
- Only a bounded set of reads is audited: exports and downloads, admin screens and sign-ins. Ordinary page views are not recorded, and command-line changes are recorded under an operating-system user, not a named person.
- Rate limits are counted per API process, not shared across replicas (the daily ceilings on the public site's stored rows are counted in the database). A refused token is counted after it has been checked. Keep volumetric protection at your ingress or web application firewall.
- A slow signing-key endpoint delays that customer's requests. Keys are fetched one provider at a time, for up to 5 seconds and at most once every 30 seconds.
- A run interrupted by a restart is not resumed. It is marked failed after a timeout and has to be started again.
- The retention purge deletes rows, not backups. Backups keep them until the backup retention has passed.
- The figure check reads figures only. It says that every amount and percentage in the text appears in the week's data. It does not check which cause or name a figure stands beside, direction words such as above or below, fractions written as words, or whether a sentence is true.
- Currency is a label, not a conversion. Amounts are printed in the tenant's currency exactly as uploaded.
- SCIM covers users and groups only. Bulk operations, sorting and password changes are not supported.
- Scheduled collection reaches public addresses only. Private endpoints and managed identity are not built; a storage account on a private network sends its files to the inbound API instead.
- Credentials and secrets for scheduled collection and webhooks are encrypted with a key derived from a deployment setting, not held in a key management service.
- The data region is a record, not a placement control. It is shown and checked against the deployment's declared region; Kelynto does not place or move data by region.
- Uploaded files are not virus-scanned. They are parsed as CSV or Excel and discarded.
- Large Excel files are expensive to read. A workbook that declares an unreasonable unpacked size is refused before it is parsed, but large files should be sent as CSV.
- Feed and signing-key addresses are checked before fetching, not pinned. Egress filtering at your network is still recommended.
- No egress firewall in the Azure templates; that stays your network's control. Private endpoints for the database, the key vault and the registry are in the private shape of the kit. A customer-managed key can be used for the database and its backups only.
- Recovery figures are design targets, and there is no round-the-clock operations cover. Alerts are emails to a mailbox.
- The incident response plan is written and has not yet been exercised.
The review pack
A written pack for security, IT and privacy teams is sent on request. It answers from the code and the deployment templates, and gives the status of every answer.
- A filled-in questionnaire in the style of CAIQ Lite and SIG Lite.
- Architecture and data flows, with diagrams.
- Subprocessors for each hosting option, as a register fixed in your contract.
- The incident response plan, backup and restore, and recovery targets.
- A shared responsibility matrix for the managed and the customer-hosted models.
Have your own questionnaire? Send it. Every answer comes back in writing, from the code and the documents.
Questions about contract documents, data processing terms and pricing are answered under Commercial on the Answers page.
Ask for the review pack, or send a security questionnaire.
Email marc@kelynto.com Trust center Deployment models