ASK · INCLUDED ON EVERY PLAN
The answer is in the code. Getting it out takes an afternoon.
Releap reads your repositories, tickets, documents and confirmations together, and answers from all four — with the file and line, so you can check it.
What people actually ask
Questions no single source can answer.
“How does suppression work now?”
The code says one thing, the wiki page says another, and the wiki page is eight months old. Releap answers from the code and flags the page.
“Why is it built this way?”
The reason is in a confirmation somebody recorded when it shipped — with their name and the date on it.
“Is this a bug or intended?”
If somebody confirmed it, there is a recorded answer. If nobody has, Releap says that instead of guessing.
“What changed here in the last quarter?”
Merges, the decisions they made, who confirmed them, and which release described them to customers.
“Which code handles refunds?”
Semantic search over what the code does, not over what it is named.
“What breaks if I change this?”
What calls it, which documentation describes it, and which customers asked for the current behavior.
Where the answer comes from
Four sources, ranked so the true one wins.
There is a row cap of 50,000 in export/filter.go:88. Marco Silva confirmed on
2 August that this is intended — Postgres times out above roughly that size.
REL-4877 makes it configurable and is targeted for 12 September.
SOURCES
• export/filter.go:88 code, current
• Confirmation · Marco Silva, 2 Aug why
• REL-4877 what happens next
• docs/exports.md describes code that has since changed
Confirmations outrank everything for “why” questions. A wiki page tells you what somebody believed in March. A confirmation tells you what a named person attested to, on a date, about code that is still there.
When it does not know
It says so, and that costs nothing.
The failure mode of every tool in this category is a confident answer assembled from thin evidence. It is worse than no answer, because somebody acts on it.
Not enough evidence
It returns what it found and says the evidence is thin. No model runs, and it does not count against your plan.
Nobody has decided
Different from “cannot find”, and said differently. The behavior exists; the intent was never recorded. That becomes a question for the team.
Out of scope
It will not answer questions about things it cannot see. It says so and points you somewhere useful.
Where you ask
In Slack, in the app, or from your own AI client.
Slack
Most questions arrive as a message to a colleague. Ask Releap in the same channel instead.
The app
Full sources, follow-up actions, and the trail from a question to the code and back.
Your coding agent, over MCP
Claude Code and any MCP client get the same context. Generation runs on your model account, so it is unlimited and free.
Everything over MCP is free, on every plan
Your agent asks; your own model account answers. Nothing about it is metered by us, so nothing about it is billed by us. It is the reason a small team can put Releap in front of a large codebase without watching a meter.
And it keeps up
Documentation that tells you when it has gone wrong.
Write a page, bind it to the code it describes, and Releap flags the section when that code changes — naming the contradiction and drafting the correction from the confirmation that caused it.
A page can go stale silently for a year. It cannot go stale silently here, which is the only reason documentation stays worth reading.