Docs / Integrations

Gemini CLI

Hydrate memory in Gemini CLI via MCP. One command to register, one prompt to verify.

Prerequisites

Both tools installed and on your $PATH.

  • Hydrate: 
    curl -fsSL gethydrate.dev/install | sh
    This places hydrate-mcp alongside the main hydrate binary.
  • Gemini CLI: see google-gemini/gemini-cli for install instructions. Version 0.26 or later.

Register Hydrate with Gemini

One command:

gemini mcp add hydrate hydrate-mcp

That writes a stdio MCP entry into your Gemini settings pointing at the hydrate-mcp binary on your $PATH. Gemini will spawn it on demand the first time a prompt triggers a tool call.

Verify the registration

gemini mcp list

Expected output:

Configured MCP servers:

  hydrate (stdio) - Disconnected

"Disconnected" is correct. Stdio MCP servers are spawned on demand — they only go "Connected" while a tool call is in flight. If you see hydrate in the list at all, registration worked.

First prompt

Paste this and send:

gemini --prompt "Call hydrate_recall with query='tech stack' and summarise what you find in 3 bullets"

Gemini will invoke the Hydrate tool, receive rendered facts from your context store, and answer from them. If you see three bullets grounded in real project facts, you're done.

Configuration

The hydrate-mcp binary reads its config from environment variables at call time:

  • HYDRATE_API — base URL for the Hydrate API (default: http://localhost:8089).
  • SESHELL_CONTEXT_USER — your Hydrate user ID (Enterprise only).
  • SESHELL_CONTEXT_KEY — your Enterprise API key.
  • SESHELL_CONTEXT_APP — your Enterprise app reference.

Free-tier users running the local Hydrate daemon can skip the SESHELL_* vars entirely — the defaults point at localhost:8089 and that's all you need.

Make recall automatic (optional)

Gemini is tool-use, not auto-inject. The model only calls hydrate_recall when it has a reason to — usually because you prompt it, or because a system instruction nudges it. If you want Hydrate to feel like it "just works" across every prompt, add a system instruction to ~/.gemini/settings.json: 

{
  "system_instruction": "At the start of any non-trivial task, call hydrate_recall with a short query describing the task to load relevant project memory. Prefer facts from the recalled context over assumptions."
}

That's the closest Gemini gets to Claude Code's auto-inject hooks. It's honest: the model still has to decide, but the instruction makes it almost always decide yes.

Tools exposed

  • hydrate_recall(query, project?) — returns rendered facts for the given query, optionally scoped to a project.
  • hydrate_save_fact(text, category?) — writes a new fact to your store.
  • hydrate_list_projects() — lists the projects tracked in your store.

Troubleshooting

  • Gemini can't find hydrate-mcp. Run which hydrate-mcp. If it's empty, re-run the install one-liner and confirm the binary landed on your $PATH. On macOS, check ~/go/bin and /usr/local/bin.
  • Enterprise user seeing empty recalls. The binary is reading localhost:8089 by default. Export HYDRATE_API, SESHELL_CONTEXT_USER, SESHELL_CONTEXT_KEY, SESHELL_CONTEXT_APP in the shell that launches Gemini CLI.
  • Gemini never calls the tool. Gemini invokes tools only when the prompt gives it a reason. Either be explicit ("call hydrate_recall..."), or add the system instruction above to your settings.