boxer — Agent Plugins package
Generated by boxer package plugin dev. Do not edit; regenerate.
One directory, valid for every client at once (Agent Plugins 1.0.0):
| Path | Standard | Purpose |
|---|---|---|
plugin.json | Agent Plugins manifest | name, version, description |
skills/boxer/ | Agent Skills | SKILL.md, scripts/ (run, task, status, brief) and references/BRIEF.md |
mcp.json | Agent Plugins MCP config | the run tool: boxer_run, boxer_status (boxer mcp) |
AGENTS.md | agents.md | the same instruction for clients that read a context file |
<reverse-domain>/ | client extension directories | hooks and shims per client; every other client must ignore them |
Client directories, each with its own README and install steps:
| Client | Directory | Native files this directory also carries |
|---|---|---|
| Claude Code | com.anthropic.claude-code/ | .claude-plugin/plugin.json, agents/boxed.md |
| Codex CLI | com.openai.codex/ | .codex-plugin/plugin.json, .agents/plugins/marketplace.json (local marketplace wrapper) |
| Grok Build | ai.x.grok/ | .grok-plugin/plugin.json |
| Gemini CLI | com.google.gemini-cli/ | gemini-extension.json (hooks copied to hooks/ in the gemini-cli view only) |
| Kimi Code | ai.moonshot.kimi-code/ | |
| DSH | com.deepseek.dsh/ | |
| OpenCode | ai.opencode/ | |
| pi | works.earendil.pi/ |
The namespace names are boxer's choice until each client publishes its own; the spec only requires
a reverse-domain form. boxer package <client> renders the subset of this directory that one
client reads (its view); boxer install <client> writes the same files into a repository's own
configuration. boxer must be on PATH; hooks and the MCP server call it by name.
Nothing here is rendered from anyone's configuration: every file is byte-identical for every user
apart from the release version. Configuration reaches the agent at run time, through boxer brief,
the session-start hooks, and the boxer://brief MCP resource.