Targets matrix

The canonical reference for what APM deploys, where, for every supported harness. Use this page to choose a target, debug an unexpected deploy location, or confirm whether a primitive is supported on a given tool.

For background on the target model, see Primitives and targets. For the runtime CLI surface, see apm targets and apm compile. For the primitive types themselves, see Primitive types.

Summary

Target	Deploy root	instructions	prompts	agents	skills	commands	hooks	mcp
copilot	.github/	[x]	[x]	[x]	[x]	[ ]	[x]	[x]
claude	.claude/	[x]	[ ]	[x]	[x]	[x]	[x]	[x]
grok-build	.grok/	[x]	[ ]	[x]	[x]	[x]	[ ]	[ ]
cursor	.cursor/	[x]	[ ]	[x]	[x]	[x]	[x]	[x]
codex	.codex/ + .agents/	[ ]	[ ]	[x]	[x]	[ ]	[x]	[x]
gemini	.gemini/	[ ]	[ ]	[ ]	[x]	[x]	[x]	[x]
antigravity	.agents/	[x]	[ ]	[ ]	[x]	[ ]	[x]	[x]
opencode	.opencode/	[ ]	[ ]	[x]	[x]	[x]	[ ]	[x]
windsurf	.windsurf/ + .agents/	[x]	[ ]	[ ]	[x]	[x]	[x]	[x]
kiro	.kiro/	[x]	[ ]	[x]	[x]	[ ]	[x]	[x]
intellij	user MCP config; files via Copilot	[x] (*)	[x] (*)	[x] (*)	[x] (*)	[ ]	[x] (*)	[x]
agent-skills	.agents/	[ ]	[ ]	[ ]	[x]	[ ]	[ ]	[ ]

Skills deploy to .agents/skills/ for Copilot, Cursor, OpenCode, Gemini, Antigravity, Codex, and Windsurf by default (see Skills convergence below). Claude, Grok Build, and Kiro keep target-native skill directories.

(*) For intellij, file primitives route through the Copilot profile: instructions, prompts, agents, and hooks use .github/, while skills use .agents/skills/. The IntelliJ-specific adapter configures MCP only.

copilot-cowork (Microsoft 365 Copilot), copilot-app (GitHub Copilot desktop App), grok-cloud (xAI Grok Cloud), openclaw (OpenClaw agent runtime), and hermes are gated behind experimental flags and not listed above. See Experimental.

Post-install instruction compilation

After a project install stages dependency instructions, the APM CLI requires a separate root-context compile for codex, gemini, and opencode, plus experimental hermes when enabled. It emits the req-tg-007 reminder for those targets. All other targets in this matrix either deploy instructions as native per-file rules, do not support dependency instructions, or have no verified root-context reader, so they do not trigger that reminder. A target not classified here does not trigger it by default.

Detection and resolution

apm install and apm compile resolve the active target list with this priority:

1. --target / --all on the command line.
2. targets: in apm.yml.
3. Auto-detection from filesystem signals (table below).

For MCP installation, the equivalent explicit legacy --runtime flag also has highest priority. MCP machine discovery runs only when targets: is omitted (or legacy all is treated as omission). Declared targets therefore produce portable MCP ownership in apm.lock.yaml; omitted targets intentionally make that ownership machine-dependent.

apm install fails closed when no target can be detected. apm compile retains its documented unsignalled fallback because install writes runtime-specific configuration while compile only generates project output. Use apm targets to preview the resolved list.

Detection signal whitelist

Target	Signals (any one activates the target)
copilot	.github/copilot-instructions.md file, or .github/instructions/, .github/agents/, .github/prompts/, or .github/hooks/ directory
claude	.claude/ directory, or CLAUDE.md file
grok-build	.grok/ directory
cursor	.cursor/ directory, or .cursorrules file
codex	.codex/ directory
gemini	.gemini/ directory, or GEMINI.md file
opencode	.opencode/ directory
windsurf	.windsurf/ directory
kiro	.kiro/ directory
intellij	Global github-copilot/intellij/ config directory (MCP runtime discovery only)

IntelliJ-specific integration is MCP-only and writes JetBrains Copilot’s user-scope mcp.json. That global signal does not auto-select file-primitive deployment. When intellij is selected explicitly, package file primitives use the Copilot profile. intellij does not participate in plain all expansion.

agent-skills is a canonical target key; antigravity is explicit-only for auto-detection. Both are available with --target and can be listed in a project’s apm.yml targets: field so contributors running plain apm install pick them up automatically.

copilot-cowork, copilot-app, grok-cloud, openclaw, and hermes are experimental targets that require apm experimental enable <name> before use. They are selected with --target only and cannot be listed in apm.yml (the canonical targets validator will reject them).

copilot

GitHub Copilot (CLI and IDE).

• Detection. .github/copilot-instructions.md.
• Deploy directory. .github/ at project scope; ~/.copilot/ at user scope.
• Supported primitives. instructions, prompts, agents, skills, hooks, mcp.
• File conventions.
  • instructions: .github/instructions/<name>.instructions.md
  • prompts: .github/prompts/<name>.prompt.md
  • agents: .github/agents/<name>.agent.md
  • skills: .agents/skills/<name>/SKILL.md
  • hooks: .github/hooks/<name>.json
  • generated: .github/copilot-instructions.md (compile output)
• User scope. Partial. prompts deploy under ~/.copilot/prompts/; instructions from all packages are concatenated into ~/.copilot/copilot-instructions.md (Copilot CLI reads only that single file at user scope). User-scope deploys land under ~/.copilot/, not ~/.github/; hook script commands are written as absolute paths so Copilot CLI can invoke them from any working directory.
• Global compile. apm compile -g can also render global instructions to ~/.copilot/AGENTS.md for root-context readers that honor AGENTS.md.

claude

Claude Code.

• Detection. .claude/ directory, or CLAUDE.md.
• Deploy directory. .claude/ (project and user scope; user scope honors CLAUDE_CONFIG_DIR if set).
• Supported primitives. instructions, agents, skills, commands, hooks, mcp. (No prompts.)
• File conventions.
  • instructions: deployed directly by apm install to .claude/rules/<name>.md
  • agents: .claude/agents/<name>.md
  • commands: .claude/commands/<name>.md
  • skills: .claude/skills/<name>/SKILL.md
  • hooks: merged into .claude/settings.json
• Compile output. CLAUDE.md; instructions already deployed under .claude/rules/ are omitted from CLAUDE.md to avoid duplicate context.

cursor

Cursor.

• Detection. .cursor/ directory, or legacy .cursorrules file.
• Deploy directory. .cursor/.
• Supported primitives. instructions, agents, skills, commands, hooks, mcp. (No prompts.)
• File conventions.
  • instructions: .cursor/rules/<name>.mdc
  • agents: .cursor/agents/<name>.md
  • commands: .cursor/commands/<name>.md
  • skills: .agents/skills/<name>/SKILL.md
  • hooks: .cursor/hooks.json
• User scope. Partial. instructions is excluded at user scope; Cursor reads global rules from its Settings UI rather than from disk.
• Global compile. apm compile -g can render global instructions to ~/.cursor/AGENTS.md for root-context readers that honor AGENTS.md; Cursor global rules still use the Settings UI.
• Caveat. Command files use the shared claude_command transformer today; Cursor-specific frontmatter keys (author, mcp, parameters, …) are dropped at install time and surfaced via diagnostics.

codex

OpenAI Codex CLI.

• Detection. .codex/ directory.
• Deploy directory. .codex/ plus .agents/ for skills.
• Supported primitives. agents, skills, hooks, mcp. (No instructions, prompts, or commands.)
• File conventions.
  • agents: .codex/agents/<name>.toml
  • skills: .agents/skills/<name>/SKILL.md
  • hooks: .codex/hooks.json
• Compile output. AGENTS.md only. Per-file instructions are not installed for Codex.

gemini

Gemini CLI.

• Detection. .gemini/ directory, or GEMINI.md.
• Deploy directory. .gemini/ (project and user scope).
• Supported primitives. commands, skills, hooks, mcp.
• File conventions.
  • commands: .gemini/commands/<name>.toml
  • skills: .agents/skills/<name>/SKILL.md
  • hooks: merged into .gemini/settings.json
• Compile output. GEMINI.md. Gemini CLI does not read per-file rules from .gemini/rules/, so instructions is compile-only.

antigravity

Google Antigravity CLI (agy), successor to Gemini CLI.

• Detection. None — explicit-only for auto-detection. Antigravity shares the cross-tool .agents/ root, so there is no unique auto-detect signal. Select it with --target antigravity or list it in apm.yml targets:; it is not part of --target all. Project-scope MCP writes are opt-in: .agents/ must already exist (APM does not create it automatically for MCP).
• Deploy directory. .agents/ (project scope); ~/.gemini/ (user scope).
• Supported primitives. instructions, skills, hooks, mcp.
• File conventions.
  • instructions: .agents/rules/<name>.md (formatted natively with trigger: glob and globs frontmatter mapped from the package applyTo patterns)
  • skills: .agents/skills/<name>/SKILL.md
  • hooks: .agents/hooks.json (Antigravity’s native schema: PreToolUse/PostToolUse/PreInvocation/PostInvocation/Stop)
  • mcp: .agents/mcp_config.json (project; mcpServers key) or ~/.gemini/config/mcp_config.json (user)
• Compile output. AGENTS.md. Supports compilation deduplication: if .agents/rules/ exists and contains at least one deployed instruction rule file (for the discovered .apm/instructions/*.instructions.md set), those instructions are omitted from AGENTS.md to avoid duplicate context.

opencode

OpenCode.

• Detection. .opencode/ directory.
• Deploy directory. .opencode/ at project scope; ~/.config/opencode/ at user scope.
• Supported primitives. agents, commands, skills, mcp.
• File conventions.
  • agents: .opencode/agents/<name>.md
  • commands: .opencode/commands/<name>.md
  • skills: .agents/skills/<name>/SKILL.md
• Caveat. OpenCode has no hooks concept; the hooks primitive is silently skipped for this target.
• Global compile. apm compile -g writes ~/.config/opencode/AGENTS.md from global instructions.

windsurf

Windsurf / Cascade.

• Detection. .windsurf/ directory.
• Deploy directory. Native primitives deploy under .windsurf/ at project scope and ~/.codeium/windsurf/ at user scope; skills converge on .agents/skills/ at both scopes (~/.agents/skills/ at user scope).
• Supported primitives. instructions, skills, commands, hooks, mcp.
• File conventions.
  • instructions: .windsurf/rules/<name>.md
  • skills: .agents/skills/<name>/SKILL.md
  • commands: .windsurf/workflows/<name>.md
  • hooks: .windsurf/hooks.json
• Agents. Not deployed. Cascade auto-invokes any SKILL.md by its description: frontmatter, so a separate agents primitive would collide with skills on the same path. Ship personas as skills under .apm/skills/<name>/SKILL.md instead.
• User scope. Partial. instructions is excluded at user scope; Windsurf stores global memory in a single ~/.codeium/windsurf/memories/global_rules.md file with a different format.

kiro

Kiro IDE/CLI v3 unified agent harness.

• Detection. .kiro/ directory.
• Deploy directory. .kiro/ (project and user scope).
• Supported primitives. agents, instructions, skills, hooks, mcp.
• File conventions.
  • agents: .kiro/agents/<relative-stem>.md — identity derives from the relative path. Only description, model, and tools frontmatter are emitted; name and unknown fields are stripped. Tools are permission-bearing: APM fails closed (no partial write) if any tool value is outside the approved set (read, write, shell, web, subagent, knowledge, context, todo_list, @mcp, @builtin, *). Kiro may warn and fall back if the specified model is unavailable; APM passes model values through without validation. Ref: kiro.dev/docs/custom-agents/ (accessed 2026-08-03).
  • instructions: .kiro/steering/<name>.md with inclusion: always or inclusion: fileMatch frontmatter
  • skills: .kiro/skills/<name>/SKILL.md
  • hooks: one JSON file per hook action under .kiro/hooks/
  • mcp: .kiro/settings/mcp.json (project) or ~/.kiro/settings/mcp.json (user)
• MCP shape. JSON mcpServers entries use command/args/env for stdio and url/headers for remote servers. Kiro resolves ${VAR} placeholders at runtime, so APM preserves them rather than writing secrets to disk.
• Scope. Covers the documented Kiro IDE and CLI v3 layout (unified harness). Ref: kiro.dev/docs/cli/v3/ (accessed 2026-08-03).

intellij

GitHub Copilot for JetBrains IDEs.

• Detection. MCP runtime discovery uses the global github-copilot/intellij/ config directory. It does not auto-select a file-primitive target.
• Deploy directory. User-scope mcp.json; see the JetBrains integration guide for OS-specific paths. macOS and Linux use $XDG_CONFIG_HOME (default ~/.config), while Windows uses %LOCALAPPDATA%.
• Supported primitives. The IntelliJ-specific adapter supports MCP. Instructions, prompts, agents, and hooks deploy through the Copilot profile under .github/; skills deploy under .agents/skills/.
• Scope. MCP configuration is user scope only. File primitives use the project or user scope selected for the Copilot profile. IntelliJ does not participate in plain all expansion.

agent-skills

Cross-client shared skills directory.

• Detection. Never auto-detected. Select with --target agent-skills.
• Deploy directory. .agents/.
• Supported primitives. skills only.
• File conventions. .agents/skills/<name>/SKILL.md.
• Use case. Author-time target for shipping a SKILL bundle that any Skills-aware client (Codex, Copilot CLI, Claude Code, etc.) can read without per-tool deployment.

grok-build

Grok Build native configuration.

• Detection. Auto-detected when .grok/ exists.
• Selection. Included in all; no experimental flag is required.
• Deploy directory. .grok/ at project scope; ~/.grok/ at user scope.
• Supported primitives. instructions, agents, commands, and skills.
• File conventions. .grok/rules/*.md, .grok/agents/*.md, .grok/commands/*.md, and .grok/skills/<name>/SKILL.md.
• Compile behavior. Produces AGENTS.md.

grok-cloud (experimental)

xAI Grok Cloud skills deployment.

• Detection. Never auto-detected. After enabling the experimental flag, selecting --target grok-cloud creates the deploy directory when needed.
• Enable. apm experimental enable grok-cloud.
• Deploy directory. .grok/ at project scope; ~/.grok/ at user scope.
• Supported primitives. skills only.
• File conventions. .grok/skills/<name>/SKILL.md.
• Compile behavior. apm compile --target grok-cloud is a successful no-op.

openclaw (experimental)

OpenClaw agent runtime.

• Detection. Never auto-detected. Select with --target openclaw after enabling the experimental flag.
• Enable. apm experimental enable openclaw.
• Deploy directory. .agents/skills/ at project scope (identical to agent-skills); ~/.openclaw/skills/ at user scope (--global).
• Supported primitives. skills only.
• File conventions. .agents/skills/<name>/SKILL.md (project) or ~/.openclaw/skills/<name>/SKILL.md (user).
• Note. At project scope the output is identical to agent-skills. The --global user path is the distinguishing capability, deploying skills where OpenClaw reads its managed/local skill directory (priority 4 in the OpenClaw loading order).

Skills convergence

Most targets with a skills primitive deploy to .agents/skills/<name>/SKILL.md. Claude, Grok Build, Kiro, and experimental Grok Cloud keep target-native skill directories.

To restore the pre-convergence per-target layout (skills land under each target’s own root), use the --legacy-skill-paths flag on apm install or set APM_LEGACY_SKILL_PATHS=1.

MCP servers

MCP is not a TargetProfile primitive; it is wired by a separate integrator that writes per-client config files (e.g. .vscode/mcp.json, .cursor/mcp.json, .claude.json, .kiro/settings/mcp.json) for every target in the active set that has an MCP client adapter. Active set follows the same --target > targets: > auto-detect chain as apm install: a runtime with an adapter but outside the active set is skipped and APM emits an [i] Skipped MCP config for X (active targets: Y) line so the gate decision is observable. The matrix above marks mcp supported when an adapter exists; whether the config gets written on a given install is a function of the active target set, not just adapter availability. See Install MCP servers for the gate behavior and apm mcp for the runtime surface.

See also

• apm targets - inspect resolved targets at runtime.
• apm compile - target selection and compile flags.
• Primitive types - what each primitive is.
• Primitives and targets - conceptual model.