Install Ricord in Gemini CLI

Prerequisites

• Google Gemini CLI installed (gemini --version works).
• Node.js ≥ 18 with npm on your PATH.
• A Ricord account — sign up if you don't have one.

Step 1 — Install the Ricord CLI

npm install -g ricord

Verify:

ricord --version
# → ricord vX.Y.Z

Step 2 — Set up Ricord

ricord setup

setup signs you in via the browser (API key lands in the OS keychain), scans your history for preferences, then auto-detects installed AI clients and wires each one. For Gemini CLI, it writes to ~/.gemini/settings.json. Gemini CLI uses the standard MCP mcpServers JSON key (same shape as Claude Desktop, different file). The added block:

{
  "mcpServers": {
    "ricord": {
      "command": "ricord",
      "args": ["mcp"]
    }
  }
}

Existing entries under mcpServers are merged, not overwritten. Verify the block landed:

cat ~/.gemini/settings.json | grep -A4 '"ricord"'

Step 3 — Start a fresh Gemini session

Gemini CLI reads settings.json at startup. If you have an active gemini session running, exit it and start a new one:

gemini

Inside the session, run /mcp to confirm Ricord appears in the loaded MCP servers list. You should see ricord listed alongside any other servers you have configured.

Step 4 — 30-second smoke test

Inside the Gemini session, type:

Remember that we deploy this project to Cloud Run via deploy.sh.

Gemini calls ricord_save. You'll see the tool call inline.

Exit the session, start a new one, type:

How does this project deploy?

Gemini calls ricord_recall and answers Cloud Run via deploy.sh. Install is good.

Workspace vs user-scope settings

Gemini CLI reads from two settings files and merges them:

• ~/.gemini/settings.json — user scope, applies everywhere. ricord install writes here by default. Recommended.
• ./.gemini/settings.json — workspace scope, applies only when you launch gemini from that directory. Use this if you want Ricord scoped to a single repo and not across all your work.

For workspace scope, copy the mcpServers.ricord block from your user settings into the repo's .gemini/settings.json, and remove it from the user file. Memory still persists to the same Ricord account — the scoping is about which sessions can read and write it.

What to expect over the next week

• Day 1: Gemini saves and recalls across gemini sessions when you ask explicitly.
• Day 3: Gemini calls ricord_get_context at session start — pulls in what it learned in previous runs without you re-explaining the repo.
• Day 7: Open the dashboard — wiki pages per repo, with the architectural decisions and deploy commands Gemini helped you figure out.

Gemini extensions and Ricord

Gemini CLI ships built-in extensions for shell, file I/O, and web search. Ricord runs alongside them — it's a separate MCP server, not an extension. If you've installed third-party extensions, none of them touch the mcpServers block. Ricord and your existing extension set coexist cleanly.

Gemini CLI auth and Ricord auth are independent

Gemini CLI authenticates to Google (OAuth or AI Studio API key). Ricord authenticates separately — handled during ricord setup. The two don't share credentials. Rotating one doesn't affect the other.

Troubleshooting

/mcp doesn't list ricord. Check that ~/.gemini/settings.json is valid JSON (a trailing comma is the most common breakage). Then fully exit Gemini and restart — settings only re-read at startup, not on file change.

ricord listed but tools don't fire. Gemini may be running with tool calls disabled. Inside the session, run /toolsto confirm Ricord's tools are enabled in the current model's tool set.

Tools loaded but recall is empty. Background indexing runs every ~10 minutes. If still empty after an hour, run ricord usage outside Gemini to verify the account is healthy.

Command-not-found: ricord (inside gemini). Gemini CLI inherits PATH from the shell that launched it. If npm's global bin dir isn't on PATH in the shell you launch Gemini from, run npm bin -g to find it and add it to your shell rc file.

Optional: build a wiki for this repo

After a few sessions, run this from any repo root to generate a structured wiki from everything Ricord has learned:

ricord build

The result appears in the dashboard — one wiki page per entity your sessions have touched.

Same setup, other clients

ricord setupauto-detects Claude Desktop, Claude Code, Cursor, Codex, and Zed if they're on the machine. Memory follows you across all six MCP clients — jump from a Gemini session to Claude Code mid-debugging, the wiki and recall stay consistent.