Discovery

Goal: a ranked list of AI agents that are getting popular, not merely popular.

Sources

Source	Tool	Signal
GitHub	gh_search (topics ai-agent, mcp-server, llm-agent, ai-agents; pushed in last 30 days; ≥20 stars)	stars, forks, homepage
Hacker News	hn_search (Algolia API; last 14 days, >5 points)	points, comments
SearXNG	searx_search (templated queries) and count_mentions	URLs, and number of distinct other domains mentioning a product

All collectors are ordinary AAA @tools running as nodes of one WorkflowGraph. Each has on_error="continue": in the first live run a source can fail (rate limit, outage) and the others still land. Signals are stored as dated rows and never overwritten, which is what makes velocity possible.

Scoring

Pure functions in discovery/scoring.py:

• pop = log1p(stars)
• vel = Δstars over 7 days / sqrt(max(stars, 50)). The square root damps giants, so a 3k-star project gaining 500 stars outranks a 100k-star one gaining 500.
• hn = log1p(points) + 0.5·log1p(comments)
• mentions = log1p(distinct other domains) (corroboration, not raw hit counts)

Each component is converted to a percentile within the current pool, weighted (default 0.30 / 0.35 / 0.15 / 0.20), multiplied by time decay (half-life 14 days) and a corroboration bonus of 1 + 0.15 × (sources − 1).

Entity resolution

A repo, its website and a Hacker News post should become one candidate.

1. Canonical keys. gh:owner/repo for repositories, web:host for product sites; aggregator hosts (Medium, Reddit, YouTube, ...) are ignored.
2. Rule merge. A repo and a site merge when the normalised names match and the host root equals the name.
3. Embedding tie-break (sqlite-vec). A repo and a site merge only if cosine similarity ≥ 0.90 and names match and their kinds differ. Two repos or two sites are different products until proven otherwise. Every merge is recorded in merges with its reason.

The bias is deliberate: prefer a duplicate over a wrong merge.

Classification

For candidates with a worthwhile score, one cheap structured call asks "is this an AI agent / MCP server / agent product, and which of 12 categories?". Output is JSON validated against a Pydantic model (GenerateTestRepair); confidence between 0.4 and 0.6 triggers a three-sample SelfConsistency vote on the label.

Run it

uv run botshub discover   # collect, score, merge, classify (about 2.5 minutes for 266 signals)
uv run botshub status     # ranked candidates

In our first live run this produced 221 candidates from 266 signals; 32 of the 40 highest-scored were classified as agents.