Add Human Approval to Gemini CLI with Impri

Add human approval to Gemini CLI: connect the Impri MCP server so risky actions wait for your one-tap approve or reject before anything runs on your machine.

Why a terminal prompt is not enough

Gemini CLI can read files, run shell commands and call tools on your behalf. The built-in confirmation prompts are useful, but they only work while you are sitting at that terminal. Start a long task, walk away, and the session either stalls or — if you pre-approved tools to keep it moving — runs unattended.

Impri adds a second kind of gate: the agent proposes an action, you get a card on your phone or in the inbox, and the agent only continues once the API reports approved. You can also edit the proposal before approving it, and every decision is kept in the audit log.

This page covers the practical setup. For the general pattern see how to add human approval to an AI agent.

Step 1 — register the Impri MCP server

Gemini CLI reads MCP servers from the mcpServers section of its settings.json. Add Impri there:

{
  "mcpServers": {
    "impri": {
      "command": "npx",
      "args": ["@impri/mcp"],
      "env": {
        "IMPRI_API_KEY": "im_your_key_here",
        "IMPRI_BASE_URL": "https://api.impri.dev"
      }
    }
  }
}

Create the key with the actions scope only — see API keys. Self-hosting? Leave IMPRI_BASE_URL at its default of http://localhost:8484 or point it at your own instance.

The server exposes three tools: impri_push_action, impri_await_decision and impri_report_result.

Step 2 — tell the agent when it must ask

The tools are only called if the agent knows to call them. Put the rule in the project context file Gemini CLI loads (GEMINI.md):

## Approval policy

Before you run any command that deletes files, pushes to a remote,
touches infrastructure, or sends anything outside this machine:

1. Call impri_push_action with kind "shell.exec", a short title, and a
   markdown preview containing the exact command and working directory.
2. Call impri_await_decision with the returned action_id.
3. Run the command only if status is "approved", and use the command
   from the returned preview — the reviewer may have edited it.
4. Call impri_report_result with "executed" or "execute_failed".

If the status is "rejected" or "expired", stop and report what happened.

A typical call from the agent then looks like this:

impri_push_action(
  kind="shell.exec",
  title="git push --force-with-lease origin release/2.4",
  preview={ format="markdown", body="```\ncd ~/work/api\ngit push --force-with-lease origin release/2.4\n```" },
  expires_in=1800,
  editable=["preview.body"]
)
impri_await_decision(action_id="act_...", timeout_s=900)

A short expires_in (the minimum is 300 seconds) suits commands: a push you approve an hour late may no longer be the push you meant.

What this does and does not guarantee

This is the part worth reading twice. A line in GEMINI.md is an instruction, and a model can ignore an instruction. Impri on its own is not a sandbox: it stores the proposed action, notifies you and holds the decision. It does not run, interpret or intercept anything.

It becomes a real gate only when the approved path is the agent's only path to the side effect:

So treat the policy file as the ergonomic layer and pair it with credential hygiene. Keep Gemini CLI's own sandbox and tool-confirmation settings on as well; Impri complements them rather than replacing them. Impri is also not a workflow engine and does not moderate content — it asks a human you trust.

Notifications for away-from-keyboard sessions

The point of an out-of-band gate is that you do not have to watch the terminal. Configure notifications (email, ntfy or web push), or approve straight from chat via Slack or Telegram. Pending cards are also listed in the inbox, where you can approve, reject or edit.

Remember that impri_await_decision blocks while it waits, so keep timeout_s modest and let the agent report a timeout instead of hanging for hours.

Next step

Get a key and push your first test action with the quickstart, then read the MCP guide for the full tool reference.