awesome-claude-code

What It Does

awesome-claude-code by hesreallyhim is the canonical community-curated directory for the Claude Code ecosystem: skills, agents, plugins, status lines, hooks, providers, observability tools, security tooling, and official Anthropic resources, organized across 18 categories. At 52,230 stars and 4,561 forks it is the highest-starred unreviewed awesome-list in GearScope's pipeline. The visible product is a single generated README, but the repo is a real software project: a 12-column CSV (THE_RESOURCES_TABLE_NEW.csv) is the single source of truth, config.yaml drives category ordering, generate_readme.py renders the README, and a Python package plus GitHub Actions automate community submissions, validation, and a live "repo ticker" that samples Claude Code projects across GitHub. It targets Claude Code developers and agent-tooling builders who want a vetted starting point rather than an exhaustive dump.

The Good

The curation is hand-written and fresh, not scraped. Every active entry carries an editorial blurb written by a human, not a copy of the repo's own description. The From Anthropic section lists 11 first-party resources (the skills repo, the building-effective-agents essay, the official plugin directory, the security-review action, the context-engineering and multi-agent research posts) with prose that explains why each matters. A link-freshness check I ran against 45 of the 135 GitHub entries returned zero dead, archived, or disabled repos. Of 157 total resources, 148 are active and 9 are kept but transparently marked inactive, with a Stale flag column that is currently unused (zero stale). For a list this large, that level of upkeep is rare.

The maintenance pipeline is real engineering, with tests. The repo ships 90 unit tests across six test files (resources, category management, README generation, ticker SVG generation, ticker data fetch, badge notification), and they pass in 0.76 seconds. generate_readme.py is reproducible: regenerating the README on a fresh clone produced a byte-identical file (zero diff against the committed README.md). IDs are opaque 8-character hex tokens minted with secrets.token_hex, deliberately decoupled from name, link, and category so a resource's identity survives edits. The resources/ package handles append, move, update, and PR-body rendering with line-oriented in-place CSV edits that preserve every other row byte-for-byte and fail closed on any multi-line value.

The submission gates are hard, automated, and anti-gaming. A GitHub Actions workflow (validate-new-issue.yml) enforces two eligibility rules before a community recommendation is even validated: a project must have at least 100 stars, or be at least 14 days old (measured from the first commit on the default branch, not repo.created_at, which closes the rename-and-transfer loophole) with signs of ongoing development. It also caps submissions at one open recommendation per author and auto-closes ineligible issues with an explanatory comment. Crucially, every API-dependent branch fails open (errors pass the submission through rather than rejecting it), so a transient GitHub outage cannot lock out a legitimate contributor. A second workflow regenerates the README, the "Recently Added" carousel SVGs, and the issue-form dropdown from the CSV on every push, committing the result back, so the published artifacts cannot drift from the data.

The ticker has a security conscience. The repo ticker that samples Claude Code projects excludes impostor repos named exactly claude-code, claudecode, or claude_code (the documented reason: malicious repos have historically distributed modified Claude Code builds). It excludes a permanently banned account and excludes the maintainer's own projects ("they are awesome enough"). A SECURITY.md and CODE_OF_CONDUCT.md are present, and the CONTRIBUTING guide is unusually honest about expectations: it states plainly that there is no formal review contract, that the list is meant to be selective, and that getting listed should never be part of a launch strategy.

The Bad

The license is CC BY-NC-ND 4.0, which is restrictive for an awesome-list. The NoDerivatives clause means you cannot redistribute a modified version of the list, and the NonCommercial clause means you cannot use it in a commercial product. Most awesome-lists are CC0 or MIT to encourage reuse. The maintainer explains the rationale in the LICENSE file (to prevent appropriation of the featured developers' labor), and the stance is defensible, but it is a real constraint: anyone who wants to fork the list, adapt its categories, or fold it into a commercial agent product cannot do so without contacting the maintainer for permission. This is the single biggest caveat for a reader evaluating reuse.

It is Claude-Code-only, with zero mentions of Hermes, OpenClaw, or Codex. A grep for hermes, openclaw, or nous across the README returns nothing. Many of the skills it lists (caveman, superpowers, fable-mode) are cross-host and install fine into Hermes or OpenClaw, but the list does not flag that, so a non-Claude-Code reader has to do the compatibility homework themselves. This is partly inherent to the name ("awesome-claude-code"), but it narrows the audience and the long-term relevance as the skills ecosystem fragments across hosts.

No releases and no changelog despite 1,521 commits and 14 contributors. The GitHub API returns zero releases; there are four tags but no semver release history to pin against or diff between versions. There is no CHANGELOG or HISTORY file. For a list that regenerates constantly and ships 1,500-plus commits, the absence of any versioned milestone or record of what changed when is a gap for anyone tracking it over time.

Inclusion is single-maintainer discretion with a large open queue. CONTRIBUTING states there is "no formal submission/review process" and that recommendations are reviewed "best-effort." The eligibility gate is strong and automated, but final inclusion is one person's editorial judgment, and 840 open issues (the submission backlog) suggests the best-effort cadence is behind demand. The Start Here category keeps five inactive guide entries rather than removing them, which is a minor tidiness issue.

Smoke Test Results

Host-based structural testing on a shallow clone (git clone --depth 1) of the repo at commit c408d9b, running on macOS 26.5.2 aarch64. The list is doc-only by its own SECURITY.md ("this repository does not export any executable software"), so the canonical smoke run is structural validation of the generator, the data, and the curation freshness.

Run A. Structural validation (tests and generator reproducibility)

$ venv/bin/python -m pytest -q
........................................................................ [ 80%]
.................. [100%]
90 passed in 0.76s ✅

$ venv/bin/python generate_readme.py
Wrote README.md (148 active entries, 18 categories)

$ git diff --stat README.md && git diff README.md | wc -c
0
README regenerates byte-identical to committed file ✅

Pass rate: 2 of 2. The full test suite passes, and the committed README is provably in sync with its source data (regeneration produces a zero-byte diff). This is the core integrity guarantee for a generated list and it holds.

Full sandbox log →

Run B. Content and schema validation (CSV, categories, counts)

$ venv/bin/python generate_readme.py
Wrote README.md (148 active entries, 18 categories) ✅

THE_RESOURCES_TABLE_NEW.csv schema (12 cols): ID, Display Name, Category,
 Sub-Category, Link, Author Name, Author Link, Active, Date Added,
 Last Checked, Description, Stale
TOTAL rows: 157 ACTIVE: 148 INACTIVE: 9 STALE-flagged: 0
Categories rendered: 18 (matches config.yaml single source of truth)
GitHub entries: 135 non-GitHub (anthropic.com / claude.com / guides): 22 ✅

Pass rate: 3 of 3. The 12-column schema parses cleanly, all 18 categories from config.yaml render, every row carries the full column set, and the active versus inactive accounting reconciles (148 active plus 9 inactive equals 157 total).

Run C. Functional verification (link freshness across the live ecosystem)

$ gh api repos/<owner>/<repo> --jq '{archived,disabled}' # 45-of-135 sample
Sampled 45 of 135 GitHub repos: 45 alive, 0 flagged ✅
(0 dead, 0 archived, 0 disabled across the sample)

Functional pass rate: 1 of 1 (link freshness). The curation has no rot in a representative one-third sample of its GitHub entries. The two live flows not exercised here are the Actions submission pipeline (needs a real issue) and the live ticker fetch (needs a GITHUB_TOKEN); both are covered by passing unit tests instead.

What the runs tell you

The engineering behind this list is stronger than most skill repos GearScope tests: tests pass, the README is provably generated from its data, the schema is sound, and the links are alive. For a doc-only curated directory, structural validation passing this cleanly is the meaningful functional bar, and it clears it. The only surfaces left unverified are the live CI submission and ticker-fetch flows, which are unit-tested but not exercised end to end here.

Setup Walkthrough

1. Clone the list to read or search it: git clone --depth 1 https://github.com/hesreallyhim/awesome-claude-code.git. The shallow clone is 5.5 MB across 56 tracked files. You do not need to install anything to use the list; it is a reference.
2. To run the maintenance tooling (regenerate the README, run tests, manage resources), create a venv and install the two runtime deps: python3 -m venv venv && venv/bin/pip install -r requirements.txt. The deps are PyYAML and requests.
3. Regenerate the README from the CSV to confirm reproducibility: venv/bin/python generate_readme.py. It prints Wrote README.md (148 active entries, 18 categories) and produces no diff against the committed file.
4. To recommend a resource, open a recommend-resource issue through the GitHub web UI form (not a free-text issue). The form enforces a 100-star or 14-day-plus-activity gate automatically.

Post-install gotcha: the license is CC BY-NC-ND 4.0, so you can read and share the list freely, but you cannot redistribute a modified version or use it commercially without the maintainer's permission. If reuse is your goal, check the LICENSE file first.

Alternatives

1. composio-community/awesome-claude-plugins - a Claude Code plugin marketplace with 25 installable in-repo plugins plus a working marketplace.json, reviewed by GearScope at 4/5. Prefer it when you want installable plugins rather than a discovery directory.
2. anthropics/skills - the official Anthropic Agent Skills repository (the SKILL.md format and example skills), reviewed by GearScope at 4/5. Prefer it for the canonical skill format and authoring examples rather than a community directory.
3. ksimback/hermes-ecosystem (Hermes Atlas) - the community-curated map of tools, skills, and integrations for Hermes Agent specifically, reviewed by GearScope at 4/5. Prefer it if your host is Hermes rather than Claude Code.