4nrry's agent plugins

Plugins and skills for AI agents. The skills follow the Agent Skills open standard, so their bodies and bundled scripts are portable across tools that speak it. The packaging in this repo — hooks included — targets Claude Code, which consumes it as a plugin marketplace; other formats may join as the collection grows.

Install (Claude Code)

Add the marketplace, then install a plugin:

/plugin marketplace add 4nrry/agent-plugins
/plugin install agent-fleet@4nrry

In Claude Desktop, paste https://github.com/4nrry/agent-plugins.git into the Add marketplace dialog.

Other agent tools

Codex (and the ChatGPT app), GitHub Copilot, Grok Build, Cursor, Kimi Code, ZCode, opencode, Muse Code, Hermes and Antigravity read plain skill folders. scripts/link_skills.py symlinks each skill into them, pointing at Claude Code's marketplace clone, so updating the marketplace updates every tool:

just link-skills global                          # agent-fleet, crew, media-agent (GLOBAL in the script)
just link-skills project ~/code/app mobile-agent  # into the repo's .agents/skills, git-excluded
just link-skills status ~/code/app

Tool	Global folder	Project folder
Codex and the ChatGPT app, Copilot CLI, Cursor, Kimi Code, ZCode, opencode	~/.agents/skills	.agents/skills
Grok Build, Muse Code	~/.agents/skills	.agents/skills, once the folder is trusted
Hermes	~/.agents/skills once listed in skills.external_dirs	.agents/skills, after hermes skills trust
Antigravity	~/.gemini/config/skills	.agents/skills
Claude Code	the plugins	the plugins, per project in .claude/settings.local.json

Hooks and MCP servers do not travel with a linked skill. For tools without Claude-format plugins, adapters/<tool>/ holds a bridge that global links into the tool's own folder; it runs the global plugins' hooks/hooks.json and mcp.json from the marketplace clone, so there is no second copy:

Tool	Hooks	MCP servers
opencode	adapters/opencode/4nrry-bridge.js, linked into ~/.config/opencode/plugins/	the same bridge
Hermes	adapters/hermes/4nrry-bridge/, linked into ~/.hermes/plugins/; once: hermes plugins enable 4nrry-bridge	once: hermes mcp add kinocut --command python3 --args ~/.claude/plugins/marketplaces/4nrry/adapters/mcp_launch.py media-agent kinocut
Kimi Code	once, in ~/.kimi-code/config.toml: a [[hooks]] entry each for UserPromptSubmit and PreToolUse running python3 ~/.claude/plugins/marketplaces/4nrry/adapters/claude_hooks.py <event>	once, in ~/.kimi-code/mcp.json: "kinocut": {"command": "python3", "args": ["~/.claude/plugins/marketplaces/4nrry/adapters/mcp_launch.py", "media-agent", "kinocut"]} (absolute path)
ZCode	once, in ~/.zcode/cli/config.json under hooks.events: a process entry each for UserPromptSubmit and PreToolUse, "command": "python3", "args": ["~/.claude/plugins/marketplaces/4nrry/adapters/claude_hooks.py", "<event>"]	once, in the same file: mcp.servers.kinocut = {"type": "stdio", "command": "python3", "args": ["~/.claude/plugins/marketplaces/4nrry/adapters/mcp_launch.py", "media-agent", "kinocut"]}
Muse Code	once, in ~/.config/muse/settings.json under hooks: a command entry each for UserPromptSubmit and PreToolUse running python3 ~/.claude/plugins/marketplaces/4nrry/adapters/claude_hooks.py <event>	once, in the same file: mcpServers.kinocut = {"type": "stdio", "command": "python3", "args": ["~/.claude/plugins/marketplaces/4nrry/adapters/mcp_launch.py", "media-agent", "kinocut"], "mode": "optional"}

adapters/mcp_launch.py PLUGIN SERVER starts a server from the plugin's mcp.json, so the version pin lives in one place. adapters/claude_hooks.py EVENT runs every matching hook for tools whose hooks already speak Claude Code's contract (stdin JSON, exit 2 blocks, stdout becomes context), as Kimi's, ZCode's and Muse's do; it maps their tool names (bash, exec_command) to Claude's. Both bridges do nothing inside a crew worker run (CREW_RUN).

Copilot, Codex and Cursor install the global plugins themselves: each ships an Agent Plugins plugin.json. Copilot can load them live from Claude Code's marketplace clone: copilot plugin marketplace add ~/.claude/plugins/marketplaces/4nrry. Grok and Cursor also load the Claude Code plugins themselves, hooks included. Codex can install this marketplace too; do not, or each skill loads twice. ZCode and Muse can install them too (ZCode with hooks, Muse without); do not, for the same reason. crew is linked too: a crew run sets CREW_RUN, and crew refuses to start a worker inside one.

Plugins

Plugin	What it is
agent-fleet	Rules for multi-agent work — model routing per phase, fan-out sizing, output schemas, and citation verification. Two hooks, because the prose alone measurably did not fire.
agents-md	Write, slim and audit a repo's AGENTS.md: what goes in, what comes out, per-tool limits (Claude Code, Antigravity, Codex), and a stdlib checker for missing scripts, dead paths, oversize files and the CLAUDE.md that hides AGENTS.md from Claude Code.
crew	Another vendor's agent CLI as an isolated subagent, from any host: ask a read-only question or hold a multi-turn conversation (agy, opencode, Codex, Grok, Hermes, Copilot, Cursor, Muse, Kimi, or Claude Code from another host), get a second opinion on a diff, or delegate one coding slice to agy, opencode, Copilot, Codex, Grok or Cursor under the worker CLI's own hooks and a mechanical gate. Tiers (cheap, mid, top) map to a model and effort per backend. The verdict never reads the CLI's exit 0 as success.
extension-audit	Audit skills, plugins and hooks against Anthropic's authoring guidance and current prompting research: a sourced checklist for judgment calls and a stdlib linter for the mechanical ones, which just check runs over this repo.
game-facts	Pins the installed game's version before the agent answers from training memory. A UserPromptSubmit hook reads Steam appmanifests and injects appid, buildid and update date — so the version never has to be typed into the prompt.
homelab	Software you maintain yourself on a Linux box, where the official docs stop: a Palworld dedicated server (the config it erases on shutdown), WireGuard through NetworkManager (the IPv6 leak that passes the IP test), and tarball installs (one versioned layout, and why a self-updating app must not live in /opt). Machine state stays out of the repo.
media-agent	Media production by an agent on a machine with an 8 GB GPU and commercial licensing, one skill per modality with rules that hold for any model: image and video generation, image retouching, voice-over at -16 LUFS, music and sound with per-item licenses, video composed as code, editing finished video, and a 3D product turntable. One hook warns where vipsthumbnail -o really writes.
mobile-agent	What breaks silently in a mobile app: the device layer (adb, emulator, UI dumps) and the bundler layer (Expo, Metro, prebuild). Six narrow PreToolUse guards that warn and never block.

Each plugin's own README carries its components, its measurements, and what those measurements do not establish.

Benchmarks

Improvement claims about plugins in this repo are comparisons between run records committed under bench/plugins/<plugin>/results/, collected per bench/PROTOCOL.md — per-agent token splits, resolved model IDs, paired arms with hash-checked prompts, and outcomes tagged by source (script measurement vs orchestrator assertion). A claim without records is marketing and does not belong here.

Records collected before that protocol existed, or by a foreign harness, are published as raw output with their hashes and are labeled imported in the first paragraph of their claims file, together with every field they lack and every claim that lack forbids.

Findings that belong to no single plugin — a fact about model routing or harness behaviour that any plugin would reuse — go in bench/shared/.

Repository checks

just check

Runs bench/validate.py: manifests carry no duplicated metadata, every hook command points at an executable file, every run record conforms to bench/schema/run.schema.json, every eval_ref resolves and hashes to the eval_sha256 beside it, every bench/ path cited in a plugin description exists, shellcheck passes, and every script advertising --self-test passes it, and the extension-audit linter reports no errors (its warnings are listed, not fatal). CI runs the same one command.

Layout

.claude-plugin/marketplace.json     entries are {name, source}; plugin.json is
                                    the single authority for plugin metadata
plugins/<plugin>/                   the installable plugin, with its own README
bench/PROTOCOL.md  validate.py      how records are collected, and the checker
bench/schema/                       run record + eval file shapes
bench/shared/                       findings belonging to no single plugin
bench/plugins/<plugin>/             evals, results and scripts for that plugin

Licence

MIT — see LICENSE.