Skip to content

minimax-ai/github-explore

v2.1.0MIT

Search, discover, summarize and audit GitHub through the gh CLI: find repos about a topic, map a field landscape, trending, repo overviews, similar projects, code search, issue/PR search and org audits — backed by 9 zero-config Python scripts with relevance ranking and quota-aware searching. Read-only by default; opt-in write management (repos/issues/PRs/workflows/secrets/releases plus auth/config and gh api mutations) requires target, impact and minimal-data disclosure plus explicit user confirmation before execution. Honours GH_HOST (defaults to github.com; routes all API traffic to e.g. *.ghe.com GitHub Enterprise Cloud or github.acme.com GitHub Enterprise Server — credentials do not cross hosts, GH_TOKEN/GITHUB_TOKEN applies to github.com+*.ghe.com only, GHES on-prem requires GH_ENTERPRISE_TOKEN, and the skill does not validate the target).

GitHub

github-explore

GitHub search, discovery, summarization and audit through the gh CLI, backed by 9 zero-config Python scripts with relevance ranking, smart deduplication and quota-aware searching.

The problem

Finding the right repositories on GitHub is noisy: broad searches drown in high-star general-purpose repos, awesome-list directories mix with real projects, and search quotas are easy to blow. github-explore packages a proven discovery workflow — dual-scope semantic search, topic mining, multi-axis field mapping with signal flags (canonical / awesome / list), trending, similar projects, code/issue search and org audits — so an agent can answer "what exists and what matters" in one skill.

Try it

Install from /plugins → Local, then ask:

find repos about vector databases written in Python

Expected result: a ranked, deduplicated list of repositories with relevance scores, filtered of forks/archived/awesome-list noise, plus a layered markdown summary (or --format json for pipelines).

Other examples:

what's trending in LLM topics this month
give me an overview of langchain-ai/langchain
find projects similar to vercel/next.js
audit the vercel org by language

Requirements

  • Python 3.10+ (scripts are zero-dependency stdlib-only)
  • gh CLI 2.x, authenticated (gh auth login)
  • Network access to your GitHub host — default api.github.com; GH_HOST redirects it (e.g. GitHub Enterprise Server, see "Data and network"; --hostname only exists on a few subcommands, not globally)

Capabilities & permissions

  • Read-only discovery (default). The 9 entry-point Python scripts (find_repos / discover / explore / trending / repo_summary / find_similar / code_search / search_issues / org_landscape) only read GitHub data. They never create, modify or delete anything.
  • Opt-in write management (requires confirmation). The skill can also drive gh management commands — create/update/delete/dispatch across repos, issues, PRs, workflows, secrets/variables, releases, org labels, gists, codespaces and SSH/GPG keys. Before any write command runs, the agent must state the exact target, the impact (including whether it is reversible), the minimal data it will change, and obtain your explicit confirmation. Read-only discovery is always the default.
  • Authentication diagnostics use gh auth status only. The skill never prints tokens: gh auth token and gh auth status --show-token are not used, and GH_TOKEN must never be echoed.

Data and network

  • All API access goes through your own authenticated gh CLI (api.github.com REST, search and GraphQL; git remotes only for clone/fork operations). Quota limits apply (search ~30/min, core ~5000/hr) and the scripts stay within them (--max-workers 2 default, retry on 403/429).
  • GH_HOST changes the network destination. It applies to all gh subcommands (git-only operations still respect git remote, not GH_HOST). The default is github.com; if GH_HOST=github.acme.com is set (GHES on-prem) or GH_HOST=acme.ghe.com (GHEC tenancy), API calls / search / GraphQL go there instead. --hostname is not a global flag — it only exists on 11 subcommands (gh auth *, gh api, gh attestation *); use GH_HOST=... gh <cmd> as the universal pattern.
  • Credentials are per-host. The hosts.yml store and keyring (gh:<host>:<user> namespace) key by host, so a PAT issued for one host will not authenticate to another (unless injected via env var — see next bullet).
  • Env var overrides follow a stricter rule: GH_TOKEN / GITHUB_TOKEN are read only for github.com / *.ghe.com / github.localhost; GHES on-prem requires GH_ENTERPRISE_TOKEN / GITHUB_ENTERPRISE_TOKEN. Set GH_TOKEN on a GHES host and gh will silently fall back to hosts.yml, not your env var.
  • The skill does not validate, restrict or warn about GH_HOST: you are responsible for confirming the target before any write command runs. Cross-host mistakes are equivalent to cross-credential leaks (issue on the wrong org, PR on the wrong repo, workflow triggered with the wrong token).
  • Permission scope = whatever scopes your gh token already has; the plugin requests none and stores none.
  • Remote side effects are limited to the opt-in write commands above; discovery scripts cause no remote changes.
  • Data flow: queries and their results pass between your machine and the configured GitHub host via the gh CLI. Markdown reports may be written to a local temp file (override with --output). No credentials are stored by this plugin; error output is redacted for credential patterns (see "Error redaction" below); no telemetry, no third-party services.

Error redaction (honest scope)

The 9 discovery scripts pipe gh errors through _lib.redact_secrets(), so credential-shaped strings (ghp_*, github_pat_*, Bearer *, token=*, GH_TOKEN=*, GITHUB_TOKEN=*) are masked before you see them.

The agent does not run the 9 scripts for every gh call — for ad-hoc management commands it invokes gh directly via the Bash tool, and that stderr is not auto-redacted. Before sharing such output, pipe it through the redactor:

gh <cmd> 2>&1 | python scripts/redact_stderr.py

redact_stderr.py reuses the same regex set. Redaction is best-effort: non-standard secret shapes can still slip through, so always scan the output before pasting into a transcript.

Security model

  • Default read-only; write operations require target / impact / minimal data disclosure and explicit user confirmation before execution.
  • Token-displaying commands are removed; authentication diagnostics only use gh auth status; error paths redact credential-shaped output.

License

MIT