Skip to content

Repository files navigation

Agent Plugins

Build, validate, and use portable Agent Plugins.

This is HiAI's curated collection of portable Agent Plugins — trusted MCP and CLI integrations paired with focused Agent Skills, safe defaults, and reproducible builds. It ships alongside the two tools that make the format work end to end:

Every plugin here is generated from a declarative plugin.yml source, byte-identical to a fresh Builder 0.0.9 regeneration, and passes the Doctor 0.0.6 validator.

Agent Plugins Builder Doctor Release License Plugins

Agent Plugins Ecosystem

Component What it is Get it
Collection (this repo) 13 curated, validated Agent Plugins with safe defaults Browse the plugins
Builder Create, migrate, and package Agent Plugins HiAi-gg/agent-plugins-builder · bunx @hiai-gg/agent-plugins-builder
Doctor Validate, diagnose, and safely fix Agent Plugins HiAi-gg/agent-plugins-doctor · bunx @hiai-gg/agent-plugins-doctor

Builder and Doctor are independent projects. They are linked here as tooling, not copied into this repository.

Plugins

The collection contains exactly 13 active plugins. Runtime status follows the vocabulary in docs/COMPATIBILITY.md; trust labels are defined in docs/TRUST_POLICY.md.

Plugin What it does Integration Status Trust
github Review pull requests, triage issues, and investigate CI and repositories. Remote MCP + Skills User auth required Trust A
agent-browser Drive a real browser: read accessible snapshots, click, fill, navigate, and test web flows. CLI + Skills Runtime verified Trust B
context7 Fetch current library and framework documentation when coding agents need up-to-date API knowledge. MCP + Skills Runtime verified Trust A
firecrawl Search, scrape, crawl, and extract structured data from the web. MCP + Skills Runtime verified Trust A
redis Inspect and diagnose Redis data, streams, and memory behavior. MCP + Skills Runtime verified Trust A
sentry Triage production errors, regressions, and performance issues. Remote MCP + Skills User auth required Trust A
supabase Inspect Supabase schemas, data access, Row Level Security, and auth. MCP + Skills Config verified Trust A
figma Bring Figma design context into agent workflows and support design-to-code tasks. Remote MCP + Skills User auth required Trust A
cloudflare Inspect and debug Workers, DNS, request handling, and Cloudflare configuration. Remote MCP + Skills User auth required Trust A
notion Search, understand, and summarize Notion workspace knowledge. Remote MCP + Skills User auth required Trust A
docker Safe Docker and Compose inspection and troubleshooting workflows. Skills-only Skills-only Trust A
kubernetes Read-first Kubernetes inspection, workload diagnosis, and cluster troubleshooting. MCP + Skills Runtime verified Trust B
postgresql Safe, read-only PostgreSQL inspection, query diagnosis, and health workflows — powered by the bundled HiAI Native MCP. Bundled MCP + Skills Runtime verified HiAI Native

Trust labels: Trust A = first-party / vendor-owned upstream (most plugins); Trust B = established community upstream that passed explicit review (agent-browser, kubernetes); HiAI Native = runtime owned and maintained directly by HiAI (postgresql's bundled MCP).

PostgreSQL: the postgresql plugin is the collection's HiAI Native highlight — its MCP server is bundled with the plugin and maintained by HiAI (no third-party PostgreSQL MCP dependency). PostgreSQL 14–18 are tested and 19 beta compatibility is tested (pre-release; not production-supported). See docs/POSTGRESQL_SUPPORT.md.

Why Agent Plugins

Agent Plugins is a portable, vendor-neutral standard for packaging Agent Skills and MCP servers so they work across compatible clients. This repository publishes HiAI's plugins in that format — one structure, reproducible from declarative sources through Agent Plugins Builder.

Every plugin ships safe defaults: read/inspect-first behavior, no credentials committed, no write-capable defaults, and no arbitrary community MCP dependencies. See docs/TRUST_POLICY.md and docs/SECURITY.md.

For Users: Installing Plugins

Each plugin directory contains both runtime files (what your Agent Plugins client loads) and authoring sources (used only to regenerate the plugin). Install the runtime files only.

Copy (runtime):

plugin.json
mcp.json                  (where present — every plugin except docker and agent-browser)
skills/*/SKILL.md         (every skill directory)
README.md
LICENSE

postgresql additionally requires its bundled MCP server (referenced by mcp.json):

packages/postgres-mcp/    (postgresql only — the bundled HiAI PostgreSQL MCP)

Do NOT copy (authoring sources / development files):

plugin.yml                (declarative authoring source)
skills-src/               (markdown skill sources)
docs/                     (development documentation)

For example, installing the PostgreSQL plugin:

# 1. copy the runtime files from plugins/postgresql into your client's plugin folder:
#    plugin.json, mcp.json, skills/, packages/postgres-mcp/, README.md, LICENSE
# 2. set DATABASE_URL for the MCP subprocess
export DATABASE_URL=postgresql://user:pass@localhost:5432/db

Most MCP-based plugins need no install step — the client launches the server from mcp.json. If your client validates with Agent Plugins Doctor, note that plugin.yml and skills-src/ produce an informational DOC-5003 ("extra files at plugin root"); they are intentionally kept in this repository as the reproducible authoring source (see below) and should not be copied into an installed plugin.

Multi-client marketplace distribution

The collection ships Git marketplace manifests for the clients that support them, plus direct-install routes for the rest. Marketplace behavior is client-specific; see docs/INSTALLATION.md for full per-client steps and docs/CLIENT_DISTRIBUTION_MATRIX.md for the acceptance matrix.

  • Codex / ChatGPT — Git marketplace at .agents/plugins/marketplace.json (source: "git-subdir" entries); register with codex plugin marketplace add HiAi-gg/agent-plugins then codex plugin add <name>@hiai-agent-plugins. Live-verified against codex-cli 0.146.0 (13-plugin discovery, 5 end-to-end installs).
  • Cursor — team marketplace manifest at .cursor-plugin/marketplace.json (relative ./plugins/<name> sources); import via Dashboard → Plugins → Team Marketplaces → Import from Repo.
  • GitHub Copilot — marketplace at .github/plugin/marketplace.json (relative sources); register with copilot plugin marketplace add HiAi-gg/agent-plugins.
  • VS CodeChat: Install Plugin From Source (paste https://github.com/HiAi-gg/agent-plugins) or the chat.plugins.marketplaces setting; VS Code reads the same .github/plugin/marketplace.json Copilot manifest. This is the Agent Plugins channel — distinct from the VS Code extension marketplace.
  • Kiro — no Git marketplace manifest exists; use Powers direct install ("Import power from GitHub"/folder) or the curated kiro.dev/powers directory. See docs/INSTALLATION.md.

The manifests are generated (never hand-edited) from the canonical plugins/<name>/plugin.json files by scripts/generate-marketplaces.ts and validated by scripts/validate-marketplaces.ts in CI. Acceptance evidence, exact commands, and the honest per-client status (VERIFIED / DOCS VERIFIED / PARTIAL) are in docs/MARKETPLACE_ACCEPTANCE.md and docs/MULTI_CLIENT_DISTRIBUTION_REPORT.md.

For Developers: Modifying Plugins

The canonical source for every plugin is:

plugins/<name>/plugin.yml
plugins/<name>/skills-src/<name>/*.md

Edit those files (plus per-plugin content such as README.md, LICENSE, docs/, and postgresql's packages/postgres-mcp), then regenerate the structural output with Agent Plugins Builder 0.0.9:

bunx @hiai-gg/agent-plugins-builder create \
  --config plugins/<name>/plugin.yml \
  --output /tmp/plugin-regen

Never hand-edit generated structural files (plugin.json, mcp.json, skills/*/SKILL.md) — they must remain reproducible from plugin.yml. See docs/DEVELOPMENT.md for the regenerate → validate → compare workflow and CONTRIBUTING.md for contribution rules.

Build & Validate

Latest release: v0.0.2.

All 13 plugins are generated from declarative plugin.yml sources through Agent Plugins Builder 0.0.9 and must remain reproducible from them. Every plugin also passes Agent Plugins Doctor 0.0.6 and independent validation against the official 1.0.0 schemas. See docs/DEVELOPMENT.md for the full workflow.

CI evidence

Every pull request and push to main runs .github/workflows/validate.yml:

  • Exactly 13 plugins — additions, removals, and renames fail CI.
  • Builder reproducibility — all 13 plugins are regenerated with Builder 0.0.9 and plugin.json, mcp.json, and every skills/*/SKILL.md must be byte-identical to the committed files.
  • Doctor 0.0.6 — the ecosystem validator runs on all 13 plugins.
  • Secret scan — secret-like literals are rejected across generated content and plugin sources.
  • Marketplace drift + validation — the three client marketplace manifests must match a fresh deterministic generation (generate-marketplaces.ts --check) and pass semantic validation (13 x 3 plugins, sources, no secrets).

The weekly .github/workflows/release-check.yml re-runs the same reproducibility and Doctor gates plus the bundled PostgreSQL MCP unit tests.

Categories

  • Developer Workflow: GitHub, Agent Browser, Docker, Kubernetes, PostgreSQL
  • Web & Research: Context7, Firecrawl
  • Data & Backend: Redis, Supabase, PostgreSQL
  • Observability: Sentry
  • Design: Figma
  • Infrastructure: Cloudflare
  • Productivity: Notion

Trust & Security

  • Trust A: first-party/vendor-owned MCP or CLI (most plugins).
  • Trust B: established community upstream that passed explicit review — currently agent-browser (vercel-labs/agent-browser CLI) and kubernetes (containers/kubernetes-mcp-server).
  • HiAI Native: runtime owned and maintained directly by HiAI — currently PostgreSQL (bundled MCP).

All plugins default to read/inspect-first behavior; credentials are never committed. See docs/TRUST_POLICY.md and docs/COMPATIBILITY.md for the full policy and per-plugin runtime evidence.

PostgreSQL Support

PostgreSQL 14–18 tested; PostgreSQL 19 beta compatibility tested. See docs/POSTGRESQL_SUPPORT.md.

Contributing

See CONTRIBUTING.md. New plugins are not added without explicit approval.

License

MIT. Individual plugins and bundled dependencies retain their own upstream licenses.

About

A curated collection of portable Agent Plugins with trusted MCP and CLI integrations, focused Agent Skills, safe defaults, and reproducible builds.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages