Agent Locations

The default registry includes Codex, Claude, DeepSeek, CodeWhale, Copilot, OpenCode, Antigravity, Grok, Kimi, OpenClaw, and chatgpt-local-coder. The installer normally requires an existing eligible agent home; absent targets are skipped. ChatGPT Local Coder is an exception: a detected runtime config directory or (on a real-system root) chatgpt-local-coder/clc CLI can qualify the target before ~/.chatgpt-local-coder exists. An approved install may create that artifact home. On real-system roots the artifact home alone is insufficient without runtime evidence.

Use this page when checking where files will be installed or why one agent was skipped. The paths below are target locations, not source locations. Canonical source content stays in this repository under canonical/ and manifest/.

Agent	Home	Skill target	Instruction file
Codex	`~/.codex`	`~/.codex/skills/<skill>/`	`~/.codex/AGENTS.md`
Claude	`~/.claude`	`~/.claude/skills/<skill>/`	`~/.claude/CLAUDE.md`
DeepSeek	`~/.deepseek`	`~/.deepseek/skills/<skill>/`	`~/.deepseek/AGENTS.md`
CodeWhale	`~/.codewhale`	`~/.codewhale/skills/<skill>/`	not modified
Copilot	`~/.copilot`	`~/.copilot/skills/<skill>/`	not modified
OpenCode	`~/.config/opencode`	`~/.config/opencode/skills/<skill>/`	`~/.config/opencode/AGENTS.md`
Antigravity	`~/.gemini/antigravity-cli`	`~/.gemini/antigravity-cli/skills/<skill>.md`, or `~/.gemini/config/skills/<skill>.md` after vendor migration	`~/.gemini/GEMINI.md`
Grok	`~/.grok`	`~/.grok/skills/<skill>/`	`~/.grok/AGENTS.md`
Kimi	`~/.kimi-code`	`~/.kimi-code/skills/<skill>/`	`~/.kimi-code/AGENTS.md`
OpenClaw	`~/.openclaw`	`~/.openclaw/skills/<skill>/`	not modified
chatgpt-local-coder	`~/.chatgpt-local-coder`	`~/.chatgpt-local-coder/skills/<skill>/`	`~/.chatgpt-local-coder/AGENTS.md`

Source of truth vs install products. Reusable skill and runtime logic is edited under canonical/ (and installed into ~/.local/share/ai-agents-skills/runtime/). Agent skill homes and OpenClaw workspace trees are install products. Do not invent durable skill behavior only under ~/.openclaw.

OpenClaw dual-route /aas adapter (workspace, not managed skill install). Messages that start with /aas may be routed to remote-bridge from a legacy OpenClaw workspace adapter. Canonical source lives under canonical/runtime/skills/remote-bridge/, but publisher use is blocked until its destination boundary is hardened. Host and workspace secrets/state remain independently owned: normal host and adapter commands do not synchronize them. The runtime carries only inert no-write stubs under the legacy sync/publisher filenames. Replacing older managed runtime copies requires an explicitly reviewed backup-and-replace upgrade with preserved recovery data; default installation does not overwrite divergent copies, and the blocked publisher cannot replace or clean up existing workspace copies. See targets/openclaw/README.md and canonical/runtime/skills/remote-bridge/publish_openclaw_adapter.py.

Optional or compatibility skill locations:

Agent	Optional location	Meaning
Codex	`~/.agents/skills`	Optional workspace/local target where supported; not the default global target.
DeepSeek	`~/.agents/skills`, `./skills`	Workspace-local locations that may shadow global DeepSeek skills.
CodeWhale	`~/.agents/skills`, compatible harness roots	Read-only compatibility roots; the installer enables CodeWhale-only discovery so `~/.codewhale/skills` remains authoritative.
Copilot	`~/.agents/skills`	Compatibility location reported but not used as the primary target.
OpenCode	`~/.claude/skills`, `~/.agents/skills`	Compatibility locations reported but not used as the primary write target.
Antigravity	`.agents/skills`, `~/.gemini/skills`	Workspace and Gemini compatibility locations reported but not used as the global write target.
Grok	`~/.claude/skills`, `~/.agents/skills`	Compatibility locations reported but not used as the primary write target.
Kimi	none additional	Primary write target is only `~/.kimi-code`.
chatgpt-local-coder	`~/.agents/skills`	Compatibility location reported but not used as the primary write target. The host also discovers `~/.claude/skills` and `~/.codex/skills` at runtime; only `~/.chatgpt-local-coder/skills` is written here.

Optional artifact-class target directories:

Agent	Personas	Templates	Commands	Tool shims
Codex	`~/.codex/agents`	`~/.codex/templates`	`~/.codex/commands`	`~/.codex/tools`
Claude	`~/.claude/agents`	`~/.claude/templates`	`~/.claude/commands`	`~/.claude/tools`
DeepSeek	`~/.deepseek/agents`	`~/.deepseek/templates`	`~/.deepseek/commands`	`~/.deepseek/tools`
CodeWhale	`~/.codewhale/agents`	`~/.codewhale/templates`	not managed	`~/.codewhale/tools`
Copilot	`~/.copilot/agents`	not supported	not supported	not supported
OpenCode	`~/.config/opencode/agents`	`~/.config/opencode/templates`	`~/.config/opencode/commands`	`~/.config/opencode/tools`
Antigravity	`<plugin-root>/agents`	`<plugin-root>/templates`	`<skills-root>/<alias>.md`	`<plugin-root>/tools`
Grok	`~/.grok/agents`	`~/.grok/templates`	`~/.grok/commands`	`~/.grok/tools`
Kimi	`~/.kimi-code/agents`	`~/.kimi-code/templates`	not supported	`~/.kimi-code/tools`
OpenClaw	not supported	not supported	not supported	not supported
chatgpt-local-coder	`~/.chatgpt-local-coder/agents`	`~/.chatgpt-local-coder/templates`	`~/.chatgpt-local-coder/commands`	`~/.chatgpt-local-coder/tools`

Rendered artifact behavior differs by agent:

Artifact	Codex	Claude	DeepSeek	CodeWhale	Copilot	OpenCode	Antigravity	Grok	Kimi	OpenClaw	chatgpt-local-coder
Skill file in auto mode	Copied regular file.	Symlink to canonical skill when supported.	Reference adapter by default.	Copied directory-layout `SKILL.md` plus support files.	Reference adapter in `~/.copilot/skills`.	Copied native `SKILL.md` plus support files.	Copied flat Markdown skill file in `~/.gemini/antigravity-cli/skills`.	Copied directory-layout `SKILL.md` plus support files in `~/.grok/skills`.	Copied directory-layout `SKILL.md` plus support files in `~/.kimi-code/skills`.	Copy-only in fake roots for eligible `SKILL.md` files.	Copied directory-layout `SKILL.md` plus support files in `~/.chatgpt-local-coder/skills`.
Persona	TOML custom-agent file.	Markdown subagent file.	Reference prompt.	Not claimed as a registered surface.	`.agent.md` custom-agent profile.	Markdown subagent file.	Plugin-scoped Markdown agent definition.	Markdown subagent file (name/description overlay).	Markdown subagent file (`name`/`description` frontmatter).	Not supported.	Markdown file in `~/.chatgpt-local-coder/agents`; inert storage, not a registered subagent.
Entrypoint alias	Reference doc under `instructions/entrypoints`.	Command file.	Reference doc under `instructions/entrypoints`.	Not managed.	Not supported by this installer target.	Command file.	Flat Markdown global skill alias.	Command file in `~/.grok/commands`.	Not supported; skills use `/skill:<name>`.	Not supported.	Reference doc under `instructions/entrypoints`.
Management notice	Managed block in `AGENTS.md`.	Managed block in `CLAUDE.md`.	Managed block in `AGENTS.md`.	Not modified.	Not supported; Copilot instruction files are not modified.	Managed block in `AGENTS.md`.	Managed block in `~/.gemini/GEMINI.md`.	Managed block in `~/.grok/AGENTS.md`.	Managed block in `~/.kimi-code/AGENTS.md`.	Not supported; OpenClaw instruction files are not modified.	Managed block in `~/.chatgpt-local-coder/AGENTS.md`.

Instruction docs target each agent’s instructions or rules directory. When any member of the four-document writing set is planned, the installer evaluates the projected hashes of the complete set. Codex, Claude, DeepSeek, OpenCode, Antigravity, Kimi, and chatgpt-local-coder receive one short managed router in their global context only while all four target documents are exact. Grok uses its native global rules/ discovery instead of a duplicate router. Selecting the complete set also refreshes only already-managed writing-policy consumer skills and safely retires an unchanged installer-managed claim-preserving-writing.md; absent or user-modified consumers are preserved. Entrypoint aliases target Claude and OpenCode commands and Antigravity global Markdown skill aliases, but Codex and DeepSeek receive reference docs under instructions/entrypoints because equivalent slash-command loading is not assumed.

Self-improvement records are workspace data, not agent-home source files. self-improving-agent writes or reviews .learnings/ entries in the current workspace, then proposes canonical changes against this repository checkout. The generated target files should be updated through plan and install after the canonical repo change has been verified.

Copilot is included in default target detection when ~/.copilot exists. Existing repository-level Copilot files under .github/ do not activate the personal Copilot target. The installer reports repository Copilot surfaces in precheck metadata, but the home-root install path writes only ~/.copilot/skills and optional ~/.copilot/agents files.

OpenCode is included in default target detection when ~/.config/opencode exists. Project .opencode/ directories are project-local and do not activate the global OpenCode target. The installer writes global OpenCode skills, rules, agents, commands, templates, instruction docs, and declared tool/plugin artifacts under ~/.config/opencode.

Antigravity is included in default target detection when ~/.gemini/antigravity-cli exists. Project .agents/ directories are workspace-local and do not activate the global Antigravity target. The installer writes flat global Markdown skills, managed global context, and the managed ai-agents-skills plugin payload under ~/.gemini/antigravity-cli/plugins/ai-agents-skills, including no-op MCP and hook scaffolds. In the table, <skills-root> is the legacy ~/.gemini/antigravity-cli/skills or validated migrated ~/.gemini/config/skills; <plugin-root> is the legacy ~/.gemini/antigravity-cli/plugins/ai-agents-skills or migrated ~/.gemini/config/plugins/ai-agents-skills. Vendor migration markers and path validation select the layout; arbitrary compatibility links are not followed. Settings remain at ~/.gemini/antigravity-cli/settings.json in both layouts.

Grok is included in default target detection when ~/.grok exists. Project .agents/ directories are workspace-local and do not activate the global Grok target. The installer copies directory-layout skills, subagents, commands, and rules/ instruction docs under ~/.grok, writes managed instruction blocks into ~/.grok/AGENTS.md, and merges a managed [compat.claude] block into ~/.grok/config.toml. The optional autoloop Stop hook installs as a discrete ~/.grok/hooks/ai-agents-skills-autoloop.json; ~/.grok/settings.json is never written. Templates and tool shims copy to ~/.grok/templates and ~/.grok/tools as inert support storage. GROK_HOME-relocated installs are unsupported.

Kimi Code is included in default target detection when ~/.kimi-code exists. The installer copies directory-layout skills under ~/.kimi-code/skills/, personas under ~/.kimi-code/agents/, inert templates/tools/instructions support storage, and managed instruction blocks into ~/.kimi-code/AGENTS.md. Copy mode is the default. The installer does not rewrite config.toml (hooks remain manual). Unattended ARL force-continue uses drive --provider kimi. Optional host-owned multi-agent panel phases use drive --panel on|auto|off (top-level CLIs such as Claude, Codex, CodeWhale, and Kimi for advice/review; the drive primary does not nest those panel calls). See docs/multi-agent-examples.md. KIMI_CODE_HOME-relocated installs are unsupported; unset that variable before real-system install. See targets/kimi/README.md.

CodeWhale is a first-class target distinct from the legacy DeepSeek target. The installer copies directory-layout skills into ~/.codewhale/skills/, uses the shared runtime root, and merges a managed [skills] block enabling scan_codewhale_only. Existing .deepseek state remains legacy compatibility data and is not reassigned to CodeWhale.

chatgpt-local-coder is included in default target detection when its dedicated installer home exists in a fake-root fixture, or, on a real-system root, when the config directory for the current platform exists (%APPDATA% on Windows, Application Support on macOS, or XDG config on Linux) or its CLI is available on PATH. A real-system artifact home alone is not runtime evidence. Runtime detection is intentionally separate from ~/.chatgpt-local-coder: that directory is the host’s declared agentHomeDir for ai-agents-skills artifacts and may be created by the first managed install. It is the local MCP coding host published as the chatgpt-local-coder npm package (clc for short), and it runs natively on Windows, Linux, and macOS. The installer copies directory-layout skills under ~/.chatgpt-local-coder/skills/, writes managed instruction blocks into ~/.chatgpt-local-coder/AGENTS.md, and copies personas, templates, instruction docs, entrypoint reference docs, and tool shims into the matching subdirectories. Copy mode is the default because native Windows symlink creation is privilege-gated. The installer merges skills.scanHostOnly=true into the host config so ~/.chatgpt-local-coder/skills is authoritative while shared and explicit roots remain available as fallbacks. The host loads ~/.chatgpt-local-coder/AGENTS.md as its user-level memory file. Its credentials live in the host config directory (~/.config/chatgpt-local-coder by XDG default; %APPDATA% on Windows and Library/Application Support on macOS), not in the agent home, so the installer never touches them. The shared writing router includes chatgpt-local-coder’s auto-loaded AGENTS.md. It routes prose tasks to the full documents on demand instead of importing them inline, which would exceed the host’s default project-memory line limit and truncate later overlays.

OpenClaw is included in default target detection when an eligible .openclaw fake-root home exists, and remains fake-root-only before native target evidence. precheck --json --agents openclaw still reports the .openclaw home shape and current gates for targeted checks, but real .openclaw writes, runtime-backed skills, support files, symlink/reference modes, and instruction-file edits remain blocked until the OpenClaw install-target plan requirements are met.

These optional artifact classes are intentionally not installed by default. They require explicit artifact selection because commands, personas, hooks, and tool shims can affect behavior more broadly than a normal skill directory.

Instruction files are modified through managed marker blocks only. Uninstall and rollback remove only recorded managed blocks and managed files; surrounding user text is outside installer ownership. The optional management-notice:repo-management artifact is also a managed block in the instruction file; it does not replace existing user instructions.