Skip to main content

Gemini CLI (Beta)

devcoach runs the full coaching loop in Google's Gemini CLI: the MCP server (tools + resources), automatic lesson cues after each task (Gemini's AfterAgent/BeforeAgent hooks mirror Claude Code's Stop/UserPromptSubmit), and the devcoach agent skill. Requires Node.js ≥ 24 and Gemini CLI with hooks support.

Beta

The Gemini CLI integration is beta: the coaching engine is the same battle-tested core used with Claude Code, but the Gemini hook wiring is newer. Please report anything odd on GitHub issues.

Option A — one command (recommended)​

npx -y devcoach install --gemini

This registers everything devcoach needs:

  • MCP server — via gemini mcp add when the gemini CLI is on your PATH (falls back to writing ~/.gemini/settings.json)
  • Hooks — AfterAgent (lesson cue after each task) + BeforeAgent (priming) in ~/.gemini/settings.json
  • Skill — the devcoach coaching skill into ~/.agents/skills/devcoach/ (the cross-tool Agent Skills directory Gemini reads natively)

Check the wiring anytime with npx -y devcoach doctor, and re-run the install after upgrading devcoach to refresh the skill.

Option B — the Gemini extension​

The devcoach extension bundles the MCP server, hooks, and skill in one package. Grab devcoach-gemini-extension-<version>.zip from the latest release, unzip it, and:

gemini extensions install <path-to-unzipped-folder>
Pick ONE option

Install the extension or run devcoach install --gemini — not both. Both register the same hooks, and a double registration would double-count your interactions (devcoach detects the extension and skips the duplicate hooks, but keep it to one on purpose). devcoach doctor warns if both are active.

How pacing works​

Identical to Claude Code: lessons are rate-limited (default max 2/day, 4h gap) and paced every N interactions per session. A hook-forced retry never re-cues (Gemini sends the same stop_hook_active guard Claude does), and plan-style turns are unaffected.

→ Next: Coaching in your agent.