Gemini CLI

Gemini CLI is Google’s open-source agentic terminal tool powered by Gemini models. devai:install configures it with a GEMINI.md project guidelines file, MCP server registration via the gemini mcp add CLI, and Marko skills distributed to .gemini/skills/.

Prerequisites

• Gemini CLI installed: npm install -g @google/gemini-cli
• Authenticated: gemini auth login (Google account or API key)
• marko/devai installed (see Installation)

What devai:install writes

Running marko devai:install with Gemini CLI detected produces the following:

GEMINI.md                          # Project guidelines for Gemini CLI
.gemini/skills/                    # Marko skill files distributed for Gemini CLI
AGENTS.md                          # Shared guidelines file (devai-managed marker block)

MCP registration is performed via the Gemini CLI (gemini mcp add) rather than by writing a config file. Gemini CLI does not implement LSP registration.

GEMINI.md

The root GEMINI.md file receives Marko project guidelines inside a <!-- BEGIN/END marko:devai --> marker block, so your own edits outside the markers survive re-runs (see Editing generated files):

• Module structure and naming conventions
• Available MCP tools and their usage patterns
• Project-specific guidelines from every installed package’s resources/ai/guidelines.md

MCP registration

The installer calls gemini mcp add -s project -t <transport> <serverName> <command> [args] to register marko mcp:serve as a project-scoped MCP server. Gemini CLI stores this registration in its own configuration; no .gemini/settings.json file is written by devai:install.

Skills

Marko skill bundles are written to .gemini/skills/ so Gemini CLI can reference them during agentic tasks.

AGENTS.md

devai:install writes the shared AGENTS.md guidelines inside a <!-- BEGIN/END marko:devai --> marker block: created if absent, and on later runs only the marked region is refreshed, so content you add outside the markers is preserved. Remove the markers to take full ownership — devai then leaves the file alone. See Editing generated files.

Manual verification

1. Open a terminal in your project root.
2. Run gemini to start an interactive session.
3. Ask: What MCP tools are registered? — Marko tools should be listed.
4. Ask: Search Marko docs for "observers" — search_docs should return results if a docs driver is installed.
5. Verify GEMINI.md exists in the project root and .gemini/skills/ contains skill files.

Agent-specific tips

• Large context: Gemini models support very large context windows. You can include more of your codebase in a single session without hitting limits. The query_database tool works well for pulling live schema data into context.
• /tools command: In an interactive Gemini CLI session, type /tools to list all registered MCP tools including the Marko ones.
• Checkpointing: Gemini CLI supports session checkpointing. When resuming a checkpoint, the MCP server reconnects automatically on first tool call.

Package READMEs

• marko/devai
• marko/mcp