How to get human approval into your coding agent

Every coding agent that can call MCP tools can be wired to pause on risky actions and wait for a human decision through Impri, with a one-line install per agent.

The pattern underneath is the same everywhere: the agent calls impri_push_action with a preview, then impri_await_decision, and only acts on approved.

Pick your agent:

• Claude Code — claude mcp add impri -e IMPRI_API_KEY=... -e IMPRI_BASE_URL=https://api.impri.dev -- npx @impri/mcp
• OpenAI Codex — codex mcp add impri --env IMPRI_API_KEY=... --env IMPRI_BASE_URL=https://api.impri.dev -- npx -y @impri/mcp
• Cursor — .cursor/mcp.json
• Windsurf — ~/.codeium/windsurf/mcp_config.json

Any other MCP-compatible client works the same way — see the MCP server reference for the full tool list and config shape, or the /agents overview for a side-by-side of all four.

The pattern, in short

1. Get an API key at impri.dev or self-host with docker compose up (see self-hosting).
2. Register the @impri/mcp server with your agent's CLI or config file (links above) — set IMPRI_BASE_URL to https://api.impri.dev for the cloud, or leave it unset for self-hosted (it defaults to http://localhost:8484).
3. Add a short instruction to your agent's system prompt or rules file: before any action with a real side effect, call impri_push_action, then impri_await_decision, and only proceed on approved.
4. Decide from the web inbox at app.impri.dev, or from your phone if you've set up notifications (Slack, Discord, Telegram, ntfy, or web push).

Where to go deeper

• The full human approval pattern — the propose → await → report loop in detail, with REST and MCP examples side by side.
• Use cases — concrete gates for overnight runs, database writes, outbound messages, deploys, refunds, and publishing.
• MCP server reference — all 8 tools, full input/output schemas.