awesome-llm-apps

What It Does

awesome-llm-apps by Shubham Saboo (Unwind AI) is a curated collection of more than 100 open-source AI agent apps, RAG tutorials, and voice and generative UI agents, with a dedicated agent_skills/ directory of portable SKILL.md skills that work across Claude Code, Codex, Cursor, Copilot, Antigravity, OpenClaw, and Hermes. At 129,471 stars it is the highest-starred agent-skill-adjacent repo GearScope has reviewed. The portable skills layer is small (5 skills plus one full-stack app), but each is built to the agentskills.io specification with real scripts, progressive disclosure references, and a five-tier eval harness that runs in CI. The broader collection is a gallery of runnable Python apps (single-agent, multi-agent, voice, RAG, MCP) that ship with their own requirements files and tutorials on Unwind AI.

The Good

The eval harness is the strongest of any community skill pack reviewed. The agent_skills/evals/ directory implements a five-tier model that explicitly follows addyosmani/agent-skills for the first three tiers, then adds two of its own because these skills ship executable code. Tier 1 runs skill_lint.py (vendored from a skill-builder skill) against the agentskills.io spec. Tier 1b runs skill_scanner.py (vendored from an agent-security-auditor skill) against the OWASP Agentic Skills Top 10. Tier 2 runs run_trigger_evals.py to check that each skill's description vocabulary routes correctly against near-miss negatives. Tier 2b runs per-skill deterministic test_*.py suites against synthetic fixtures. Tier 3 is on-demand behavioral evals in Anthropic's skill-creator schema. All of it is stdlib-only Python, runs in CI on every pull request touching agent_skills/**, and the evals README names the two real bugs it caught before merge (a macOS symlink-path bug that silently disabled relapse detection, and a filesystem-ordering bug that mis-attributed a kill-chain).

The skills pass every tier clean on a fresh clone. I ran the full CI stack on the host against the cloned repo. skill_lint.py --strict returned 0 errors and 0 warnings on all 5 skills. skill_scanner.py returned 0 CRITICAL, 0 WARN, 0 INFO across the whole tree. run_trigger_evals.py returned all clear for 5 skills (weakest positive score 1.00 vs strongest near-miss 0.00). The three deterministic test suites returned 18 of 18, 16 of 16, and 15 of 15 checks (49 of 49 total). That is a perfect run across every free tier.

The scripts work on real input, not just fixtures. I ran commit-archaeologist/scripts/archaeologist.py against the repo's own README and got valid JSON with an introducing commit, a one-entry timeline, authorship, and an intent signal. The output was byte-identical across two runs (deterministic). A bad path returned error: file is not tracked at HEAD: nonexistent.py with no traceback. I ran scope-creep-detector/scripts/scope_creep.py against a synthetic diff containing a new package.json dependency, and it correctly produced a SCOPE REPORT flagging the dependency as "likely creep" with a "SPLIT OR JUSTIFY" recommendation. Both scripts are pure stdlib with no third-party imports.

The skills are written by someone who understands the failure modes. commit-archaeologist ships an explicit Evidence Rules section that warns against blame-ownership fallacies, co-change correlation treated as dependency proof, and "temporary" markers assumed to still apply. project-graveyard redacts author names in its kill-chain reports and tracks relapse on every resurrection it prescribes. thinking-out-loud quarantines the model's own inferences away from the user's phrasing so the user verifies what the model believes, not just what it doubts. advisor-orchestrator-worker budgets a three-tier model team so a run cannot burn a hole in an API bill. These are concrete, named failure modes, not generic advice.

The privacy and security stance is explicit and verifiable. Every skill declares whether it makes network calls (none of the five do). The agent_skills README states "nothing asks your agent to curl pipe bash anything, ever" and instructs users to read every SKILL.md and script before installing. The security scanner is vendored into CI rather than left as an afterthought.

The Bad

The headline 100-plus conflates full apps with portable skills. The README says "100-plus open-source AI agents, agent skills, and RAG apps," but the actual portable agent skills (folders with a SKILL.md that installs into an agent) number 5 in the agent_skills/ directory, plus 3 more buried in generative_ui_agents/ai-mcp-app-builder/. The remaining 90-plus entries are standalone Python applications (RAG tutorials, voice agent teams, multi-agent fraud investigators) that each ship their own requirements file and run as apps, not as drop-in skills. A reader who comes for "agent skills" and clones 168 MB expecting 100 installable skills gets a smaller skills layer than the marketing implies. The agent_skills README is honest that it is "more coming, released one at a time," but the top-level framing sets a different expectation.

No releases, no tags, and no version discipline at the repo level. Despite 129,471 stars and 99-plus contributors, the GitHub API returns empty arrays for both releases and tags. Per-skill versioning is inconsistent: thinking-out-loud declares v1.3.0 while the other four sit at v1.0.0, with no changelog explaining the gap. There is no CONTRIBUTING guide and no CHANGELOG anywhere in the tree. For a repo of this scale, the absence of any release artifact or contribution guide is a real gap.

Two of five skills are doc-only inside a framework that says scripts must earn their place. The agent_skills README sets a bar: "Most skills on registries are text-only prompt dumps. Skills here have to earn their place: real scripts, deterministic work runs as code." advisor-orchestrator-worker and thinking-out-loud both ship zero scripts and rely on references alone. They are well-written references, and advisor-orchestrator-worker carries three reference files (the most of any skill here), but they do not meet the repo's own stated bar for executable content. The self-improving-agent-skills entry is not a portable skill at all: it is a full Next.js frontend plus Python backend app that requires a Gemini API key to run.

The clone is heavy with no documented sparse path. A --depth 1 clone is 168 MB (the GitHub API reports 215 MB), inflated by gallery images under docs/. A user who only wants one skill still downloads the whole collection. The install instructions point at npx skills add for a single skill, which avoids the full clone, but the manual-copy path does not document a sparse-checkout alternative.

Smoke Test Results

Host-based testing on a shallow clone (git clone --depth 1) of the repo at commit 9f1f80a, running the full eval stack plus functional scripts on macOS aarch64. All eval tools are vendored Python 3 stdlib with no dependencies.

Run A. Structural validation (skill_lint.py against the agentskills.io spec)

$ python3 agent_skills/evals/tools/skill_lint.py agent_skills/commit-archaeologist --strict
Linting .../agent_skills/commit-archaeologist (strict=on)
 PASS: 0 error(s), 0 warning(s) ✅

$ python3 agent_skills/evals/tools/skill_lint.py agent_skills/project-graveyard --strict
 PASS: 0 error(s), 0 warning(s) ✅

$ python3 agent_skills/evals/tools/skill_lint.py agent_skills/scope-creep-detector --strict
 PASS: 0 error(s), 0 warning(s) ✅

$ python3 agent_skills/evals/tools/skill_lint.py agent_skills/advisor-orchestrator-worker --strict
 PASS: 0 error(s), 0 warning(s) ✅

$ python3 agent_skills/evals/tools/skill_lint.py agent_skills/thinking-out-loud --strict
 PASS: 0 error(s), 0 warning(s) ✅

Pass rate: 5 of 5. Every skill's frontmatter parses, names match their directories, descriptions are within limits, body sizes are under the 400-line warning threshold (the largest is project-graveyard at 229 lines), and all relative file references resolve on disk.

Run B. Security scan and trigger routing

$ python3 agent_skills/evals/tools/skill_scanner.py agent_skills
agent-skill security scan v1.0.0: 5 skill(s) under agent_skills
 advisor-orchestrator-worker: clean: no findings ✅
 commit-archaeologist: clean: no findings ✅
 project-graveyard: clean: no findings ✅
 scope-creep-detector: clean: no findings ✅
 thinking-out-loud: clean: no findings ✅
Summary: 0 CRITICAL, 0 WARN, 0 INFO ✅

$ python3 agent_skills/evals/tools/run_trigger_evals.py
PASS advisor-orchestrator-worker: 3 positives clear 3 near-misses ✅
PASS commit-archaeologist: 3 positives clear 3 near-misses ✅
PASS project-graveyard: 3 positives clear 3 near-misses ✅
PASS scope-creep-detector: 3 positives clear 3 near-misses ✅
PASS thinking-out-loud: 4 positives clear 3 near-misses ✅
trigger & routing: all clear (5 skills) ✅

$ python3 agent_skills/evals/commit-archaeologist/test_archaeologist.py
PASS: 18/18 checks ✅
$ python3 agent_skills/evals/project-graveyard/test_graveyard.py
PASS: 16/16 checks ✅
$ python3 agent_skills/evals/scope-creep-detector/test_scope_creep.py
PASS: 15/15 checks ✅

Pass rate: 8 of 8. Zero security findings across the OWASP Agentic Skills Top 10 patterns. Every skill routes its trigger vocabulary above its near-miss negatives. All 49 deterministic script checks pass.

Run C. Functional verification (scripts on real input)

$ python3 agent_skills/commit-archaeologist/scripts/archaeologist.py /tmp/awesome-llm-apps-r1 README.md --json
exit code: 0
introduced_by hash: 9f1f80a584f8
introduced_by subject: Merge pull request #1022 from thejesh23/fix/single-agent-apps-batch
timeline count: 1
authors count: 1
intent_signals count: 1
output deterministic across two runs: yes ✅

$ python3 agent_skills/commit-archaeologist/scripts/archaeologist.py /tmp/awesome-llm-apps-r1 nonexistent.py --json
error: file is not tracked at HEAD: nonexistent.py ✅ (clean rejection)

$ python3 agent_skills/scope-creep-detector/scripts/scope_creep.py --diff /tmp/test.diff
SCOPE REPORT
intent: main
files: 3 additions: 4 deletions: 0
subsystems: (root)=2, src=1
likely creep: 3
 SPLIT OR JUSTIFY: package.json (new dependency) ✅
 SPLIT OR JUSTIFY: README.md (intent has no usable keywords)
 SPLIT OR JUSTIFY: src/auth.py (intent has no usable keywords)

Functional pass rate: 2 of 2 skills with scripts verified on real input. commit-archaeologist produces correct, deterministic JSON from real git history and rejects invalid input cleanly. scope-creep-detector correctly flags a new dependency and a formatting-only file in a synthetic diff.

What the runs tell you

The eval harness is the real product here. Every free tier passes clean on a fresh clone, the two script-backed skills produce correct output on real input, and the security scanner finds nothing. The skills layer is small (5 portable skills) but each one is held to a standard most community skill packs do not meet. The only untested surface is the self-improving-agent-skills full-stack app, which needs a Gemini key and was not started.

Setup Walkthrough

1. Install a single skill with the skills CLI (auto-detects your agent): npx skills add https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills/commit-archaeologist. This drops the skill into ~/.hermes/skills/ (or the matching directory for Claude Code, Codex, Cursor, Copilot, Antigravity, OpenClaw).
2. To run the eval harness first, clone the repo and run the CI stack: git clone --depth 1 https://github.com/Shubhamsaboo/awesome-llm-apps then python3 agent_skills/evals/tools/skill_lint.py agent_skills/commit-archaeologist --strict.
3. For the script-backed skills (commit-archaeologist, project-graveyard, scope-creep-detector), no dependencies are needed. Each script is pure Python 3 stdlib. Run with python3 scripts/.py from inside the skill folder.
4. The doc-only skills (advisor-orchestrator-worker, thinking-out-loud) have no scripts to run. They load their references on demand when an agent invokes them.

Post-install gotcha: the skills install cleanly but the broader repo clone is 168 MB. Use npx skills add for a single skill to avoid the full download.

Alternatives

1. addyosmani/agent-skills - a pure engineering skills pack with the eval model this repo's tier 1 to 2 is based on. Prefer it for a larger catalog of script-backed skills with no app-gallery baggage.
2. anthropics/skills - the official Anthropic skill pack including skill-creator, whose eval schema this repo's tier 3 uses verbatim. Prefer it for the canonical authoring workflow.
3. obra/superpowers - a larger agentic skills framework with a subagent-driven methodology. Prefer it if you want an opinionated end-to-end development workflow rather than discrete engineering tools.