One URL, and your agent can read the Atmosphere
Add aturi.to/api/mcp to Claude, Cursor, or any MCP client. Your agent can then read any account’s repository, trace who references what across every app, and follow Jetstream as records land. No key, no account, nothing to install.
38 read-only tools · no API key
https://aturi.to/api/mcpWhat has @aturi.to been posting about, and which post landed best?
resolve_identity→ did:plc:6teuh…get_author_feed→ 8 posts + countsSpaces, record editing, and new client support. The Spaces announcement did best:
at://did:plc:6teuhlkizzebk6wdp42633el/app.bsky.feed.post/3mtkpzxkh5k2eaturi.to/profile/aturi.to/app.bsky.feed.post/3mtkpzxkh5k2eSettings › Connectors › Add custom connector https://aturi.to/api/mcp
claude mcp add --transport http atmosphere https://aturi.to/api/mcp
{ "atmosphere": { "url": "https://aturi.to/api/mcp" } }Add it in one line
Any client that speaks Streamable HTTP works. Clients that only speak stdio can bridge with npx mcp-remote.
Thirty-eight tools
Identity and repositories on any personal data server, the network graph, the Bluesky social layer, feeds and lists, lexicon activity, a live Jetstream tap, and the protocol docs themselves. Most of it reaches past Bluesky into the rest of atproto, which is where the questions get interesting.
- Who links to this post, from any app?
- What parameters does getAuthorFeed take?
- How do record keys work in atproto?
- It reads. No tool can post, like, follow, or edit anything, and the server holds no credentials that could.
- Beta means tool names and result shapes can still change. Nothing should pin to them yet.
- Answers come from live public services: Bluesky’s AppView, plc.directory, Jetstream, and microcosm’s Constellation, Slingshot and UFOs. When one is down or rate-limiting, the tool says so rather than guessing.
- Documentation answers are read from the atproto.com and docs.bsky.app sources at the moment you ask, and every one carries the page URL. Check the page before acting on anything load-bearing.
- Bluesky’s post search refuses requests from data-centre networks, so
search_postscan fail where every other tool works. - Posts and records are written by strangers. Treat what comes back as data to read, not as instructions to follow.
- There is no uptime promise and no support queue. One person maintains this, so be reasonable about volume.
Building software instead of prompting an agent? The same answers are plain GET endpoints, typed in the OpenAPI document and explained in the developer docs.