Security
Security and data handling
Where your data is stored, who can reach it, and what a connected coding agent can and cannot do. What we do not have yet is here too.
What broke today?
Ask about your errors and your users. The answer comes from your own data in Vinktar.
What a reviewer checks first· 8
Each answer links to the detail further down.
- Where is our data stored?In Helsinki, Finland, on UpCloud. There is one region.EU
- Does an AI model read our data?Not one of ours. Vinktar runs no model. What your own agent reads goes to its provider.No model
- Can another workspace see ours?Membership is checked on every request, and every warehouse read is bound to your ids.Kept apart
- What can the key in our app do?Send events and errors. It reads nothing.Send only
- How does a coding agent get in?Through OAuth that a person approves, for one workspace and the projects they pick.Approved
- Can an agent delete our data?Not events, errors, issues or projects. A board it deletes comes back for 30 days.No
- Is error text safe to hand an agent?It arrives fenced and marked as data. That lowers the risk of prompt injection, not to zero.Lowered
- SOC 2, SSO or a choice of region?Not yet. If your review needs one of them, we are not the right choice today.Not yet
Your data
Stored in Helsinki
The application, the event warehouse and the database run on UpCloud in Helsinki, Finland.
Where it is stored
UpCloud, Helsinki, one region. Where your data lives has what is stored where and what is backed up.
Model providers
Vinktar runs no model, so no AI company is a subprocessor. What your coding agent reads goes to that agent’s provider, under your agreement with them.
IP addresses
Reduced to a keyed, non-reversible hash before it is stored. The country is looked up at the edge, and only the country. Geolocation by DB-IP, CC BY 4.0.
Deletion
Deleting a project or a workspace purges its events and errors from the warehouse.
How long it is kept
By plan. History past the window is trimmed nightly.
| Plan | Events and issue history | Error occurrences |
|---|---|---|
| Free | 30 days | 30 days |
| Starter | 1 year | 90 days |
| Pro | 1 year | 90 days |
| Business | 2 years | 180 days |
Third parties
Stripe handles billing, Google handles sign-in if you choose it, and Resend delivers email. Each is on the subprocessor list, and a new one is added there before it first touches your data.
Access
Kept apart
What keeps one workspace out of another, and what each key can do.
Keys for project web
The two kinds a project has
vnk_pk_8f2c41d07a9ePublic by design: it ships in your app. It can only send, and it reads nothing.
vnk_sk_••••••••••••A secret. Shown once, when a person creates it, and never returned to an agent.
Between workspaces
The API checks that you are a member before it answers. Every query carries a workspace condition that cannot be removed, and every warehouse read is bound to your workspace and project ids.
SQL you type
SQL in Explore, or from an agent through run_sql, goes through our own compiler: SELECT only, over your own events, every literal bound. It runs on a read-only connection with caps on time, rows and memory.
Signing in
Sign in by email, password or Google.
Coding agents
Approved by a person
It connects with OAuth 2.1 and PKCE, so there is no API key to paste into it. A person signs in, in their own browser, and approves one workspace and the projects it may reach.
Connect Claude Code
It will read your data as you, within what you choose here.
- Workspace
- Harrowby
- Projects it can reach
- web, api
What it may do
Reading is always on. Build and configure are offered only where the person approving can make those changes themselves.
What it can do· 5
- Read events, errors, boards and trends in the scope you chose
- Define events and write the project’s notes
- Build, edit and delete boards and panels. A deleted one can be restored for 30 days
- Create alerts that reach your owners and admins
- Resolve, ignore, reopen or assign an issue
What it cannot do· 6
- Delete events, errors, issues or projects
- Create projects or API keys
- Add a webhook, Slack channel or email list
- Hide or un-hide an event or property
- Touch members or billing
- Run code or deploy anything
Checked on every call
The grant is checked against every id the agent passes, and the person’s membership by the same code the web app uses. An agent never sees more than the person who approved it.
Alerts it creates
They reach your existing owners and admins. A request that names a webhook, a Slack URL or an email list is refused with a sentence the agent can relay to you.
Hidden properties
Mark a property hidden from agents and it is refused in every query, filter and breakdown, and stripped from results. Only a person can clear the flag.
The log
Every tool call is logged with its arguments, how long it took and what it cost, on the AI agents page for 30 days. Boards and alerts an agent made are listed there with a delete for each.
Limits
Calls are budgeted per connection, per person and per workspace, at half the read budget, and every warehouse read has a deadline. A monthly allowance stops a loop. It is never billed.
Switching it off
Disconnect your own connections at any time. An admin can disconnect anyone’s, or switch agents off for the workspace, which refuses the next call. Connect your agent
Text your users wrote
Error text, as data
Error messages, event names, payloads and URLs come from anyone who can load your app, because the key that sends them is public. Some of that text will one day be written to steer an agent.
Marked as untrusted data
Captured text arrives inside a <vinktar-data> block, cleaned of anything that could close it early, and the server’s instructions say what that block is. Every link is built from ids we issued, never from a URL in your data.
What that does not fix
It lowers the risk of prompt injection. It does not remove it: an agent can still be misled, and it has other tools than ours. Check what it proposes, most of all on connections that can make changes.
Not yet
What we do not have yet
Vinktar is small, deliberately. Weigh that as the risk it is. If your review needs one of these, we are not the right choice today.
SOC 2
Not yetNo SOC 2 report and no other certification. A reviewer gets this page, the subprocessor list and an answer by email from the person who built Vinktar.
SSO or SAML
Not yetNeither exists. Sign in by email, password or Google.
A choice of region
Not yetThere is one region: your events and errors are stored in Helsinki.
Report
Report a vulnerability
Write to john@vinktar.com
The person who built Vinktar reads it and answers.
Say what you found, the steps to reproduce it, and which workspace or project it touches if you know. The same address is in /.well-known/security.txt.