Ask about your code, not just the PRs about it.
Connecting GitHub or GitLab indexes the source itself, not only issues and review threads. “How does our auth middleware work?” gets an answer grounded in the actual implementation, cited file by file.
On connect, every repo the account can read is walked via the tree API and its text files land in the same ingest pipeline as documents: chunked, embedded, and citable from Ask, Search, and the MCP filesystem tools. Indexing starts at backfill time and a 4-hourly re-index picks up new commits; unchanged files are skipped by blob hash, so re-walks are cheap.
A deny-list runs before any file is fetched: .env files, anything that looks like secrets, credentials, or key material, node_modulesand other vendored dependencies, binary assets, and any file over 100 KB. Blocked files are never downloaded, so they can’t leak through a later layer.
Up to 5,000 files per repo are indexed, in tree order. Access mirrors the repo: code documents carry ResourceAcl rows for the collaborators, so a private repo’s code is only retrievable by people who can already read that repo. Manage both connectors at Admin → Connectors.