One understanding of your system
The Runbook: your languages, how to build and test, the data model, where the logs are, and what “production” means for you. It records what it couldn’t get running, and why. You correct it, and every job loads it.
Answers: an AI SRE that starts from a customer’s question
Most AI SREs start from an alert; Winnfield starts from what support escalates. It reads four sources together:
- The code: what the system does.
- The production data: what happened to this customer.
- The logs: what happened at the time.
- The docs: what it should do.
Every answer ends in a verdict
- Bug. A ticket with the record ids, the causal chain with every hop cited, and the repositories in scope.
- Configuration. The exact setting for the exact customer, and nothing reaches engineering. It’s checked against the docs first.
- Missing feature. A story with a plan and acceptance criteria for each repository. Product questions are asked, never guessed.
- Needs more info. Nothing filed. It says what it checked, what was missing and what would settle it.
The evidence gate: a grep is navigation, not evidence
A reported bug is dropped unless the run shows three things:
- It read the records it cites.
- It opened every repository it names.
- Its file:line resolves.
Fixes and features: an AI software engineer that tests against your records
Trusted code with no AI in it fetches the records the ticket cited. A verifier holding no production, cloud or push credential must demonstrate the fix against them. The rules:
- A run with nothing executed is downgraded.
- A run with signs of tampering is thrown away.
- When the incident can’t be reproduced, the ticket says so.
Features are demonstrated against their acceptance criteria in staging, on a disposable database, with real-world side effects forbidden.
Testing: AI QA on every fix
It runs your own suites the way the Runbook records them. Pull requests use your template, with a box ticked only where the pipeline guarantees it, and the verifier’s verdict in the notes.
Nightly: bugs found before they’re reported
The same investigation runs unprompted over your production activity, and files only what clears the evidence gate.
In Slack: an AI agent for every team
Three ways in: an @mention in a watched channel, a dedicated channel, or a DM. A late @mention reads the thread above. The answer comes first; the evidence is one click away.
In Jira: an agent on the work item
- How work starts. A work item’s Agent picker, an assignment, a workflow step or an automation sends it to work. An @mention asks a question.
- What it needs. Jira Cloud on Standard, Premium or Enterprise, with Rovo on. Each invocation uses your Rovo credits.
Included in select plans
The support desk
Add Winnfield’s chat widget to your product, and your signed-in customers get answers where they already are.
01
The customer asks in the widget.
02
Tier 1 answers from your docs, citing them, or says it’s checking.
03
Winnfield investigates and leaves an internal note for tier 1 and your team.
04
Tier 1 answers the customer with what Winnfield found.
05
Your team can reply, or file a bug with a click.
The assistant never touches your code or data. Winnfield does that, read-only, and briefs it.
Trust, earned in stages
Each stage stays off until the one before has a week of clean numbers, and each has its own blast radius.
Stage 1
Shadow
Two weeks of private answers your team rates. Blast radius: no change to your code or data.
Stage 2
Answer
Replies where your teams ask, read-only. Blast radius: no change to your code or data.
Stage 3
File
Bugs and stories with their evidence. Blast radius: a ticket someone reads.
Stage 4
Build
Fixes, tests and features on feature branches. Blast radius: a pull request your engineers review.
When it’s wrong
- 👎 asks what was wrong: the facts, the conclusion or missing context.
- The correction can become a line in your Product Brief.
- A wrong ticket can be withdrawn from the same thread.
What you never do
No prompt writing, no rules, no knowledge base, no training data, no schema documentation, no integration engineering.
Questions
Does Winnfield write to our production database?
No. Winnfield reads production through a role your DBA creates with read-only grants. We generate the exact grant statement for your database engine and measure the credential’s privileges when you connect it. Read-only is enforced by your database, not by a prompt.
Can it push to main or merge?
No. Every code change arrives as a feature branch and a pull request that your engineers review and merge. Winnfield has no merge capability, and the credential that pushes a branch never enters the environment where the AI works.
What does it need from us?
Slack; your code on GitHub; Jira Cloud on Standard, Premium or Enterprise with Rovo turned on, or Linear; a read-only production database role and a disposable staging database; a read-only, log-scoped credential for your cloud; read-only access to your docs; and test-mode keys for the third-party services your code calls. The production credential never shares a machine with your running code.
Why not just use Claude with MCP servers?
You can connect a general assistant to your systems. What MCP gives an assistant is access. It doesn’t give you a reason to trust the answer: nothing checks that a record it cites was actually read, nothing stops a configuration verdict that skipped your docs, nothing tests a fix against the records the ticket cited, and nothing proves beforehand that it can install and run your system. Those gates, the Runbook, and an architecture where production credentials never share a machine with your running code are the product.
What happens when it isn’t sure, or is wrong?
When it isn’t sure it says so: a needs-more-info answer files nothing and lists what was checked, what was missing, and what would settle it. When it’s wrong, “Not right” in the thread asks what was wrong, the correction can go into your Product Brief in one click, and a ticket filed on a wrong verdict can be withdrawn from the same thread.
How is it priced?
Prices aren’t published yet. Pilot terms are on request.