Awesome AGY Subagents

The largest open-source collection of AGY-compatible subagents and tools.

AGY is the Antigravity CLI (agy) — Google's terminal AI coding agent. An AGY subagent is a single Markdown file whose body is a system prompt and whose YAML frontmatter carries runtime configuration. This repository aggregates, converts, validates and categorizes subagents from hundreds of public repositories so they can be used immediately in AGY (and any frontmatter-subagent harness).

4,250 converted subagents · 176 contributing repositories · 152 documented tools · 602 repositories discovered (100% evaluated)

---

Quick start

1. Copy any agent into your AGY project:

   mkdir -p .gemini/agents
   cp agents/security/security-auditor.md .gemini/agents/

2. Invoke it from agy by its name:

   > ask the security-auditor subagent to review src/auth.py
   

3. Or install the entire collection at once (every agent is already validated and categorized under agents/<category>/):

   cp -r agents/* .gemini/agents/

Every agent is ready to use — no conversion, renaming, or manual work needed.

---

Repository layout

agents/           # 4,250 converted AGY subagents, organized by category
  backend/ frontend/ security/ devops/ ai/ mobile/ testing/
  research/ writing/ architecture/ ... (27 categories)
tools/            # 152 documented tools (20 core AGY tools + 132 MCP integrations)
  core/  mcp/
workflows/        # Multi-agent workflow definitions
examples/         # Composition / usage examples
templates/        # Authoring templates for new agents & plugins
docs/             # Specification, importing guide, compatibility model
validation/       # JSON Schema for the AGY format + validation results
scripts/          # discover, process_all, convert, finalize, classify,
                  # validate, dedupe, report, export
metadata/         # Machine-readable registries (agents.json, summary.json,
                  # repo_status.json, ...)
reports/          # Generated human-readable reports (summary, coverage,
                  # unsupported, compatibility, tools, mcp, ...)
exports/          # agents.csv export (agents.json is git-ignored, it is a
                  # duplicate of metadata/agents.json)
imports/          # Streamed source clones (git-ignored; deleted after scan)

---

Every agent carries

Each file in agents/ is a complete AGY subagent with:

• Name — a unique, invocation-ready slug
• Description — how the parent agent decides when to call it
• Category + tags
• Original repository, author, license, source URL, and file path
• Required / optional tools and required MCP servers
• AGY compatibility score and status
• The converted AGY system prompt
• Validation status and import timestamp

See docs/AGY-SPEC.md for the full format, and metadata/agents.json for the machine-readable registry.

---

Statistics

Generated by scripts/report.py — see reports/summary.md.

Metric	Value
Repositories discovered	602
Repositories evaluated (coverage)	602 (100%)
Agents discovered (pre-dedupe)	7,457
Duplicates removed	3,207
Agents converted	4,250
Fully AGY-compatible	3,024
Requires MCP	285
Needs tool mapping	594
Requires manual conversion	347
Conversion success rate	71.2%
Average compatibility score	91.4
Documented tools	152
Categories	27
Contributing repositories	176

Repository classification (full detail in reports/coverage.md):

Status	Repositories
Imported	170
Duplicate	6
Unsupported	375
Non-agent	50
Requires manual review	1

Top contributing repositories (full list in reports/repositories.md):

• Spielewoy/autoprompt-skill — 421 agents
• jeremylongshore/claude-code-plugins-plus-skills — 394 agents
• davepoon/buildwithclaude — 343 agents
• ruvnet/ruflo — 325 agents
• leamas-ai/leamas.sh — 319 agents
• msitarzewski/agency-agents — 271 agents
• …and many more

---

Importing & maintenance

The pipeline is fully re-runnable and streamed (clones are deleted after each repo is scanned, keeping the workspace and git history small). To refresh:

# Use the project virtualenv (PyYAML is required)
python3 -m venv .venv && .venv/bin/pip install PyYAML

# Full pipeline (discover → process → finalize → classify → validate →
# report → export)
.venv/bin/python scripts/run_all.py

# Or individual stages
.venv/bin/python scripts/run_all.py --discover --process

• Discovery queries the GitHub search API (keyword, filename:, path: and topic: searches) and writes metadata/discovered_repos.json.
• Process (process_all.py) shallow-clones every discovered repository, scans it recursively, classifies it (Imported / Duplicate / Unsupported / Empty / Non-agent / Requires manual review), extracts reusable agents into metadata/raw_records.json, and deletes the clone.
• Finalize deduplicates the raw records and writes the canonical agents/<category>/<name>.md files plus metadata/agents.json.
• Classify assigns each repository's final status and emits the coverage and unsupported reports.
• Validate checks every agent against the spec.
• Report / Export regenerate summary stats, the tool catalog and exports/.

To contribute a new upstream repository, add it to KNOWN_REPOS in scripts/discover.py and re-run scripts/run_all.py. See docs/importing.md and CONTRIBUTING.md.

---

License

Collection and tooling: MIT (see LICENSE). Individual imported agents retain their upstream licenses, recorded in each agent's agy.sources[].license field.