# divepool

> Discover Bluesky accounts and posts by topic. Semantic search across the AT Protocol network, powered by embeddings, keyword analysis, and clustering. Agent access is MCP-first.

## MCP (the agent surface)

- [MCP server](https://divepool.com/mcp): streamable-HTTP MCP. Two discovery tools: `discover` (one seed — text query, account, record, or account set (handles or DIDs) — returning either whole entities across Bluesky accounts + podcast episodes + Hacker News comments on one relative score scale, or individual records with optional topic clustering) and `inspect` (per-account topic medoids + 0-100 similarity scores for accounts you name). OAuth-gated: sign in at `/api/v1/oauth/login` ("Sign in with Bluesky"); connectors can self-onboard via the OAuth AS (PKCE + dynamic client registration).
- The former public REST endpoints (`/api/v1/search`, `/similar-accounts`, `/medoids`, `/embed`, `/embeddings`) were retired 2026-07 in favor of MCP and now return 410.

## Background

- [Homepage](https://divepool.com/): one-screen overview and topic-search entry point.
- [Talk: Keywords vs Embeddings (ATmosphereConf 2026)](https://divepool.com/presentation/PRESENTATION.html): slides on how divepool uses embeddings, c-TF-IDF, and clustering for ATproto-based social media discovery.
- [Churn dashboard](https://divepool.com/churn): feed liveness signal for Bluesky accounts.
- [Churn methodology](https://divepool.com/churn/methodology): how the dashboard answers "Is Bluesky dying?". Definitions, the per-user vs global calculation, verdict thresholds, and notes on a sampling-bias fix.
- [About](https://divepool.com/about): how divepool works — embeddings, keyword analysis, and clustering applied to AT Protocol posts. Covers what each user-facing surface returns (handle vs. topic), the discovery feed on Bluesky, and a brief note on who builds it.

## Analyses

- [Analyses index](https://divepool.com/analyses/): interactive, self-contained experiments on public AT Protocol data.
- [City Atlas](https://divepool.com/analyses/city_atlas.html): the world's ~34,000 largest cities embedded by name, laid out in 2D with UMAP and shown side by side with real geography, across seven embedding models — how much of the world map do name-embeddings recover?