QUESTIONS AND ANSWERS
The questions people ask before they trust this.
Grouped by what you are actually worried about. Pricing questions live beside the numbers on the pricing page, so there is only ever one answer to keep true.
16 questions
Your data
4 questions
- What actually leaves our network?
- A redacted version of the error payload, sent to whichever model provider is configured. Tokens, passwords, card numbers, and similar patterns are stripped before the request is built, not after the response comes back. If the answer needs to be nothing at all, run the model locally — then no incident context leaves the network in the first place.
- Which model provider is used?
- Whichever one you choose. Anything speaking the OpenAI chat-completions API works, including a model running on hardware you control. There is no provider you are obliged to use.
- Are you the controller or the processor?
- We act as your processor: you decide what is in scope and why, and we act on your instructions. The data-processing terms set out what that obliges us to do, and what it does not permit.Read the data processing terms
- How long is anything kept?
- Incidents, analyses, and the approval record stay for as long as the workspace holds them, because a decision trail with a gap in it is not a decision trail. Retention is a setting, and the period is yours to choose.
What it can and cannot do
4 questions
- Can it change something in production on its own?
- No. Every applied change waits for a person to approve it, and only travels a path you connected on purpose. Nothing is switched on by default.
- What does a proposed change look like?
- Either a pull request on a new branch — created from your default branch and opened against it, never pushed to it — or a file change prepared for a target you configured, with the backup taken beforehand kept alongside it so a revert has something to restore.
- What if the analysis is wrong?
- Assume it sometimes will be, and read the evidence rather than the conclusion. The risk score is deliberately not the model's opinion alone: a rule-based score is calculated independently from environment, severity, frequency, affected users, and the words in the error, and the final score is whichever of the two is higher. A model that underestimates a payment failure cannot talk the score down.See how the evidence is presented
- Does it need access to our code?
- Not to start — analysis works from the error payload alone. A repository connection is only needed if you want a prepared pull request, and it is scoped to the repositories you pick.
Running it
4 questions
- How quickly does anything happen?
- Issues are pulled from your error tracker every five minutes, and each registered project is checked every minute by its own uptime monitor. Analysis runs on a queue, so a burst of incidents lengthens the queue rather than dropping work.
- Do we have to change our application?
- No. There is no agent, no SDK, and no code change. The workspace reads from the error tracker your application already reports to, using a read token you create.
- Which error trackers work?
- Any Sentry-compatible tracker. You supply the instance URL, an organisation, a project, and a token, and issues start syncing on the next scheduled pass.See every connection
- Will this page us for everything?
- Only for the risk levels you allow, per project, through the channels you enable — email, Discord, or Telegram. A cooldown stops the same issue from arriving repeatedly. The default is high and critical only, and most teams tighten it further in the first week.
Getting started
4 questions
- What do you need from us to begin?
- A decision about scope — which applications and which environments — and a read token for the error tracker. That is the whole list. The first sync usually surfaces things that have been failing quietly for months.See the onboarding steps
- Can we run it ourselves instead?
- Not today. We operate it, which is what keeps the model provider, the queues, and the upgrade path our problem rather than yours. If running it inside your own network is a requirement rather than a preference, say so in an enquiry — it is a conversation, not a self-serve option.Ask about it
- Can we start with one application?
- That is the recommended way. Start with the service that breaks most often, and add the rest once it has proved itself on that one.
- How do people on our side get accounts?
- By invitation from an administrator on your side. There is no open sign-up, which is deliberate: an account in the workspace can see your incidents.
STILL UNANSWERED
If your question is not here, it is probably the interesting one.
Send it. A person reads it, and the answer usually names what we would watch and what we would leave alone.
- A scoping call takes about thirty minutes
- No agent to install and no change to your application
- Nothing is connected until you approve it