agent-commands-tool

Registers every agent as a /agent:<name> slash command. Typing /agent:planner implement user authentication directly spawns the planner agent as a subprocess with that task.

No tool call overhead, the agent runs immediately as a child process with full access to its configured tools and context.

1 Agent command picker showing available agents after typing /agent

Agent directories

Agents are loaded from multiple locations with last-wins priority:

1. Bundled agents — shipped with tallow
2. Package agents — from settings.json packages
3. ~/.claude/agents/ — user-level, Claude Code compatibility
4. ~/.tallow/agents/ — user-level
5. Project .claude/agents/ — project-level Claude Code
6. Project .tallow/agents/ — project-level (highest priority)

.tallow/ takes precedence over .claude/ at each scope level. See claude-bridge for details.

Frontmatter fields

Agents support the same frontmatter fields as subagent-tool:

• disallowedTools — tool denylist, filters against the built-in set (read, bash, edit, write, grep, find, ls)
• maxTurns — turn budget hint injected into the system prompt. Soft enforcement only — unlike subagent-tool, this extension does not kill the process at the limit because it doesn’t use --mode json for event-based turn counting.
• mcpServers — restrict MCP server access via PI_MCP_SERVERS env var

---
name: safe-reviewer
description: Read-only code reviewer
disallowedTools: bash, write, edit
mcpServers: github
maxTurns: 20
---