How Pithelo works

Every morning at 08:00 Athens time, Pithelo publishes the 5 agent-related GitHub repositories that are rising fastest. Nothing is hand-picked; the same rules run every day.

Data sources

Candidates are found with the GitHub search API (topics such as ai-agents,llm-agent, agentic-ai, mcp-server, claude-code, plus "ai agent" in names and descriptions). Star, fork, commit and licence numbers come from the GitHub GraphQL API. We record one snapshot per repository per day, and compute "trending" from those snapshots.

What counts as agent-related

A language model answers two questions about each new repository: is it primarily a tool, framework, server, skill or resource for building or running AI agents (a probability), and how useful is it to a developer building agents (a 1–5 score). Code applies fixed thresholds to the answers:

• Probability at least 0.85 and usefulness at least 2: accepted.
• Probability from 0.50 to below 0.85: held back for manual review, never picked automatically.
• Probability below 0.50, or usefulness below 2: rejected.

The ranking

Each accepted repository with at least 50 stars is scored against the others that day:

• 50%: stars gained over the last 7 days.
• 30%: growth rate, which is stars gained divided by stars a week ago, capped at 5.0 (500%).
• 20%: commits to the default branch in the last 30 days.

Each signal is turned into a percentile rank among that day's eligible repositories, and the weighted sum becomes a 0–100 score. Ties go to the more useful repository, then to the one with more new stars.

A repository needs a snapshot from exactly 7 days earlier, so a missed day can never turn "7 days" into 8. A repository picked in the last 14 days is skipped, and spam or bought-star repositories can be blocklisted by hand.

Warm-up

Until enough history exists, picks are labelled "Warming up: ranked by stars per day since creation" and only consider repositories created in the last 90 days.

README token cost

For each pick we send the full README to Anthropic's token-counting endpoint and show the result as "README ≈ N tokens to load into context". It is a measurement, not an estimate. If counting fails, the figure is left out.

Summaries

Summaries are written by an AI model from the repository's own text only, and are marked as AI-written. If the model's answer is unusable we show the repository's own description instead.

Licences

We show the licence GitHub reports (SPDX identifier) and do not filter by it. "No licence declared" means GitHub found none, which usually means all rights are reserved: check before you reuse code.

Reliability

A day is published all at once or not at all. If a run fails, the site keeps showing the previous edition. See status for the last 14 runs.