Supported AI agents

Plugin

The plugin targets AI coding agents that implement both Agent Skills and MCP as open standards. Skills, built-in agents, and MCP tools are shared across hosts. Only hooks and plugin manifests are host-specific.

Support matrix

Host	Status	Install	MCP wiring	Hooks API
Claude Code	Production	Plugin marketplace	Auto — plugin ships .mcp.json (uses archcore from PATH)	SessionStart, PreToolUse, PostToolUse
Cursor	Implemented	Plugin marketplace / local dir	User-registered from the docs/cursor.mcp.example.json template (scoped via --project ${workspaceFolder})	sessionStart, preToolUse, afterMCPExecution
Codex CLI	Implemented	codex plugin marketplace add	Auto — plugin ships .codex.mcp.json (uses archcore from PATH)	SessionStart, PreToolUse, PostToolUse (gated)
GitHub Copilot	Planned	—	—	—

Production means the plugin is deployed and used day-to-day; bug fixes and new skills ship through the marketplace.

Implemented means every skill and agent works, and all documented hooks fire, but some host-level edge cases still surface occasionally. On Codex CLI, hooks are additionally gated by a Codex-side feature flag — see Codex CLI below.

Host-specific differences

The differences are at the edges:

Plugin manifest

Aspect	Claude Code	Cursor	Codex CLI
Manifest file	.claude-plugin/plugin.json (minimal)	.cursor-plugin/plugin.json (explicit)	.codex-plugin/plugin.json (explicit)
Field pointers	Convention-based	Explicit skills, agents, hooks	Explicit commands, agents, hooks, mcpServers

MCP server registration

Aspect	Claude Code	Cursor	Codex CLI
Who registers MCP	The plugin, via .mcp.json at plugin root	The user, via ~/.cursor/mcp.json or .cursor/mcp.json (copy the docs/cursor.mcp.example.json template)	The plugin, via .codex.mcp.json referenced from manifest’s mcpServers
Command points at	archcore mcp (from PATH)	archcore mcp --project ${workspaceFolder} (from PATH)	archcore mcp (from PATH; Codex inherits shell PATH)
CLI install step	Required — install globally via archcore.ai/install.sh	Required — install globally via archcore.ai/install.sh	Required — install globally via archcore.ai/install.sh

Hook files

Aspect	Claude Code	Cursor	Codex CLI
Hook file	hooks/hooks.json	hooks/cursor.hooks.json	hooks/codex.hooks.json
Trigger names	SessionStart, PreToolUse, PostToolUse	sessionStart, preToolUse, afterMCPExecution	SessionStart, PreToolUse, PostToolUse
Write matcher	`Write	Edit`	Write only
Output protocol	Exit codes + structured JSON	JSON wrapper objects	Exit codes + structured JSON
Path env var	${CLAUDE_PLUGIN_ROOT}	${CURSOR_PLUGIN_ROOT}	${PLUGIN_ROOT} (host-neutral)

All three hook files live in the same plugin package and are selected automatically by the host.

Cursor: afterMCPExecution

Cursor exposes an extra afterMCPExecution trigger that fires after MCP tool calls succeed. The plugin uses it to run validate-archcore, the cascade staleness check, and the precision-warnings check in a single hook. On Claude Code and Codex CLI, the same work happens via PostToolUse with multiple matchers.

Codex CLI

Codex CLI uses the same PostToolUse semantics as Claude Code, plus a third write surface — Codex’s apply_patch tool — which is added to the PreToolUse matcher. Codex hooks are gated by a Codex-side feature flag:

codex features enable plugin_hooks

This flag is currently under development in Codex (default off). Without it, hooks are inert but skills, MCP, and agents still work normally.

The path env var is ${PLUGIN_ROOT} — Codex’s canonical, host-neutral name. There is no ${CODEX_PLUGIN_ROOT} variable.

Command surfaces

Skills are surfaced differently depending on host conventions:

• Claude Code and Cursor load skills directly from skills/<name>/SKILL.md.
• Codex CLI surfaces skills through thin commands/<name>.md wrappers (7 of them) — one per command.

The skill content itself is the same; only the entry point differs.

Agents

Both archcore-assistant and archcore-auditor ship as .md (Claude Code, Cursor) and .toml (Codex CLI). The TOML variants translate the same tool whitelists into Codex’s sandbox model — read-only for the auditor, workspace-write for the assistant. See Built-in agents.

Choosing a host

Use Claude Code if:

• You want the most mature path — it has the deepest testing coverage.
• You’re already invested in the Claude Code CLI workflow.

Use Cursor if:

• You prefer a full editor over a terminal-first agent.
• You want afterMCPExecution semantics (single post-MCP hook).

Use Codex CLI if:

• You want a terminal-first agent with auto-registered MCP (no manual wiring).
• You’re comfortable enabling the plugin_hooks feature flag to get full hook coverage.

All three are first-class. Skills, agents, hooks, and guardrails behave identically.

Unsupported hosts

The plugin is not available on GitHub Copilot yet, or on any other MCP-aware CLI without a plugin runtime. Those agents work via the CLI path only — MCP access and agent integrations, but no plugin skills or tracks.

Track planned host support in the plugin’s repository.

Next steps

• How plugin works — the hook lifecycle that differs per host.
• Install — marketplace commands per host.
• Troubleshooting — host-specific issues.