Understanding AT Protocol repositories
Repositories, collections, record keys, CIDs and blobs: what the explorer reads, how requests are protected and how to export your account’s data.
Repositories and collections
An account’s Personal Data Server (PDS) holds its public repository. Collections group records by schema NSID: app.bsky.feed.post contains posts, app.bsky.graph.follow contains follows. An unknown NSID belongs to another schema, not necessarily Bluesky.
The repository explorer accepts a handle, DID or AT URI. The repository view lists collections; a collection shows 50 records per page in either order. URLs use q, cursor and reverse=1, so navigation also works without JavaScript.
Keys, CIDs and references
An AT URI names the repository, collection and record key. A CID identifies a particular version of the content. Many record keys are TIDs encoding a time, but other keys, including self, have no TID date. See the AT URI converter and TID decoder.
The JSON tree escapes text and is limited to 256 KiB and depth 20. Internal AT references link to records; DIDs link to Whois. Blob references show CID, MIME type and size. A blob download is an explicit external link to the PDS; arbitrary PDS images are never loaded automatically.
Public reads and safety
The tool resolves a DID document, checks relay status and asks the PDS for com.atproto.repo.describeRepo, com.atproto.repo.listRecords or com.atproto.repo.getRecord. Every PDS request uses the SSRF guard, HTTPS, a four-second timeout, no redirects and body caps of 1 MiB, or 256 KiB for one record. Private addresses are refused. Results are cached for five minutes, collection pages for one minute. The tool cannot write.
Export and open locally
In Bluesky, choose Settings → Account → Export my data. Developers can download com.atproto.sync.getRepo?did=… directly from their PDS. Open the resulting .car in the CAR explorer. It runs in a local browser worker, with no upload, API, MCP interface or file-action analytics. Follow a reference only when you want to navigate elsewhere.
CAR v1 exports are limited to 512 MiB, 250,000 records, 500,000 sections including duplicates and 2,000 collections. Headers are capped at 64 KiB and blocks at 1 MiB before their bodies are collected. A separate 512 MiB budget accounts for retained block bytes and reconstructed index metadata. The worker reads 64 KiB slices, hashes each distinct block once and checks duplicate bodies for equality. Each local operation has a 60-second deadline; cancel a read to reuse the already loaded worker without a network request. Collections and records paginate at 100 entries. Available device memory can impose a smaller practical limit. A present commit signature is shown as unverified. Search and JSON Lines remain local; ZIP and bulk blob export are not included.