What reaches each tool¶
Skills reach every tool. The rest depends on what each tool can load.
At a glance¶
| Claude Code | Codex | OpenCode | Antigravity CLI | Cursor | |
|---|---|---|---|---|---|
| Skills | ✅ | ✅ | ✅ | ✅ | ✅ |
| Rules | ✅ CLAUDE.md |
✅ AGENTS.md |
✅ AGENTS.md |
✅ AGENTS.md |
↪ plugin only |
| Commands | ✅ | ✅ prompts | ✅ | ➖ skills are its slash commands | ⏳ |
| Agents | ✅ subagents | ↪ role-switch fallback | ✅ subagents (rendered) | ↪ fallback | ↪ fallback |
| Guard hooks | ✅ allow, ask, deny | ❌ rules only | ✅ plugin, deny only | ⏳ rules only | ✅ shell and reads, deny only |
| MCP servers | ✅ | ✅ | ✅ | ⏳ | ⏳ |
✅ supported · ↪ partly, or another way · ⏳ not yet · ➖ not needed · ❌ not possible today
"Fallback" means the tool has no named subagents, so Gatework switches roles inside one session.
Where each piece goes¶
One row per tool: where apply puts each kind of content, and what a repo's brainrc.toml
writes for it (team setup). A dash means that tool gets none of it. Paths are the
defaults; the variables in the last column move them (configuration).
This table is generated from harnesses.toml, and a test fails when the two
disagree. The table at the top is written by hand: when you change harnesses.toml, update it too.
| Tool | Skills | Rules | Agents | Commands | Guard hooks | MCP servers | In a repo | Moves with |
|---|---|---|---|---|---|---|---|---|
| Claude Code | ~/.claude/skills |
~/.claude/CLAUDE.md |
~/.claude/agents |
~/.claude/commands |
~/.claude/settings.json: PreToolUse, SessionStart |
~/.claude.json |
.agents/skills/, AGENTS.md, .claude/skills/ (copy), CLAUDE.md (imports AGENTS.md) |
CLAUDE_CONFIG_DIR |
| Codex | ~/.codex/skills |
~/.codex/AGENTS.md |
— | ~/.codex/prompts |
— | ~/.codex/config.toml |
.agents/skills/, AGENTS.md |
CODEX_HOME |
| OpenCode | ~/.config/opencode/skills |
~/.config/opencode/AGENTS.md |
~/.config/opencode/agents (rendered) |
~/.config/opencode/command |
plugin ~/.config/opencode/plugins/brainrc-guard.js |
~/.config/opencode/opencode.json |
.agents/skills/, AGENTS.md |
XDG_CONFIG_HOME |
| Antigravity CLI | ~/.gemini/config/skills |
~/.gemini/config/AGENTS.md |
— | — | — | — | .agents/skills/, AGENTS.md |
— |
| Gemini CLI (off by default) | ~/.gemini/skills |
~/.gemini/GEMINI.md |
— | — | — | — | .agents/skills/, AGENTS.md |
— |
| Cursor | ~/.cursor/skills |
— | — | — | ~/.cursor/hooks.json: beforeReadFile, beforeShellExecution |
— | .agents/skills/, AGENTS.md |
— |
What the tables can't show:
- Rules reach Cursor only through its plugin, because Cursor keeps its user rules in its Settings, not in a file.
- Agents: a tool with no agents path gets no agents from
apply. OpenCode can't read Claude Code's agent format (atools: Read, Grepline makes its whole config invalid), soapplylinks copies rendered to~/.config/brainrc/opencode/agents/. Each copy is a subagent denied every tool the agent lacks; itsmodelis dropped, so it uses its caller's model, andmcp__*tools aren't mapped. - Guard hooks: Claude Code can allow, ask or deny; OpenCode's plugin and Cursor's hooks can only deny, so whatever would ask is blocked (guardrails). Codex and Antigravity get the rules alone.
- Off by default: Gemini CLI was replaced by Antigravity CLI. Enable it in
harnesses.local.tomlif your account still runs it. In a repo, an off-by-default tool gets files only whenbrainrc.tomlnames it inharnesses. - Antigravity CLI's paths (
~/.gemini/config/) come from its changelog and user reports, not a live run.
Tool paths change often. To add or move a tool, edit harnesses.toml (and, for a new tool, its
display name in brainrc/capabilities.py), then run
python3 scripts/brainrc.py capabilities --write.