Search GitHub for AI Tools

AI/ML tooling moves fast. A library that was best-in-class twelve months ago is often broken against current model APIs, SDK versions, or provider behaviour today. This skill applies a stricter recency filter than the general

search-repos

skill.

When to use

Trigger phrases include:

• "Find an AI/ML library for X"
• "Is there an MCP server for Y?"
• "What's a good LLM framework for Z?"
• "Search for agent frameworks / prompt management / eval tools / RAG libraries"

Procedure

1. Frame the search with a recency cutoff

Default cutoff: last push within the last 6 months. Unless the user overrides, exclude anything staler than that at the query level.

# Compute a 6-month-ago date
CUTOFF=$(date -d '6 months ago' +%Y-%m-%d)

gh api -X GET search/repositories \
  -f q="<ai-keywords> pushed:>$CUTOFF" \
  -f sort=stars -f order=desc --jq '.items[] | {
    full_name, stars: .stargazers_count, pushed_at, updated_at,
    language, license: .license.spdx_id,
    open_issues: .open_issues_count,
    description, url: .html_url
  }'

Useful query operators to combine:

• topic:llm

  ,

  topic:agents

  ,

  topic:rag

  ,

  topic:mcp

  ,

  topic:prompt-engineering

• stars:>50

  to strip hobby projects

• archived:false

2. Tier the results

Sort candidates into three tiers based on recency, then by stars within each tier:

• Active — pushed within last 30 days
• Maintained — pushed 1–6 months ago
• Stale (⚠️) — pushed 6–12 months ago; include only if the user explicitly asked for a complete picture
• Dead — pushed > 12 months ago; exclude entirely unless specifically requested

3. Look for activity signals beyond last-push

• Recent releases —

  gh api repos/<owner>/<repo>/releases --jq '.[0].published_at'

• Commit cadence —

  gh api repos/<owner>/<repo>/commits --jq '[.[].commit.author.date] | .[:5]'

• Open PRs — healthy projects have a steady stream, not hundreds untouched

Flag projects where last-push is recent but commits are only dependency bumps (Dependabot churn) — that's not real maintenance.

4. Present results

Output grouped by tier. For each candidate include:

• Full name, stars, last push ("N days/weeks/months ago")
• Language, license, open-issue count
• One-sentence summary of what it does
• A short note on maintenance signal ("active releases", "only Dependabot commits", "solo maintainer", etc.) where visible

5. Offer next steps

• Deeper evaluation via

  evaluate-candidates

• Documentation via

  document-findings

  (AI research especially benefits from date stamping — what's good today may be obsolete in 3 months)

Guidance

• A 2k-star AI project last updated 14 months ago is almost always worse than a 200-star project updated yesterday. Say so explicitly.
• Watch for abandoned demos — viral projects from blog posts or tweet threads often look popular but have no ongoing maintenance.
• LLM API compatibility rots fast. Anything predating the current model generations may be broken.
• Do not invent projects. Report only what the CLI returned.