Impri

Impri

Human-in-the-loop approval inbox for AI agents.

Your agent wants to send the email, post the reply, run the migration. You want to see it before it happens, from your phone if you're not at a desk. Impri is the inbox in between.

Agent (Claude Code / any MCP client)  →  Impri inbox  →  You approve, edit, or reject
        ↑                                                          |
        └──────────────── agent executes, reports result ──────────┘

Watch as MP4

The problem

Agents that only read and draft are safe by construction. The moment one can send an email, post a reply, or run a write against a database, "ask the user first" becomes a system-prompt instruction — and a instruction is something a capable model can reason its way past, forget under load, or misjudge on an edge case you didn't anticipate.

Impri turns that instruction into a data dependency instead. The agent pushes a proposed action to an inbox and polls for a decision; the execution branch is only reachable once the API actually returns status: "approved". There's nothing to talk the model past — the gate lives outside the model, not inside its prompt.

60-second install

No install at all — connect straight to the hosted endpoint if your client supports a remote MCP server:

claude mcp add --transport http impri https://api.impri.dev/mcp --header "Authorization: Bearer $IMPRI_API_KEY"

codex mcp add impri --url https://api.impri.dev/mcp
# then add to ~/.codex/config.toml under [mcp_servers.impri]:
#   bearer_token_env_var = "IMPRI_API_KEY"

Or run the local package (needed if you're pointing at a self-hosted server instead of the cloud):

Claude Code:

claude mcp add impri --env IMPRI_API_KEY=im_... --env IMPRI_BASE_URL=https://api.impri.dev -- npx -y @impri/mcp

Codex:

codex mcp add impri --env IMPRI_API_KEY=im_... --env IMPRI_BASE_URL=https://api.impri.dev -- npx -y @impri/mcp

Cursor — add to .cursor/mcp.json:

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

Windsurf — add the same block to mcp_config.json:

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

Get an im_... key with one curl call — no signup form:

curl -s -X POST https://api.impri.dev/v1/signup -H "Content-Type: application/json" -d '{"name":"my-agent"}'

IMPRI_BASE_URL defaults to http://localhost:8484 (self-host) if you omit it — set it to https://api.impri.dev for the hosted cloud, as above. Self-hosting instead of the cloud? See Self-host or cloud below.

What the agent can do

The loop

1. impri_push_action(kind, title, preview)        → { action_id, status: "pending", inbox_url }
2. impri_await_decision(action_id)                 → waits for a human to approve/reject/edit, or times out
3. If approved: execute the real action
4. impri_report_result(action_id, "executed" | "execute_failed")

Or in plain REST, the same three calls:

ACTION=$(curl -s -X POST https://api.impri.dev/v1/actions \
  -H "Authorization: Bearer $IMPRI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"kind":"email.send","title":"Outreach: Acme","preview":{"format":"markdown","body":"Hi Sarah, ..."}}')
ID=$(echo "$ACTION" | jq -r .id)

# poll (or long-poll) until a human decides
DECISION=$(curl -s "https://api.impri.dev/v1/actions/$ID" -H "Authorization: Bearer $IMPRI_API_KEY")

# only now, and only if approved
[ "$(echo "$DECISION" | jq -r .status)" = "approved" ] && send_email "$(echo "$DECISION" | jq -r '.decision.final_preview.body // .preview.body')"

The human doesn't have to be staring at a dashboard: an action can notify Slack, Discord, Telegram, ntfy, email, or a generic webhook, and the reviewer can edit the draft before approving — the agent receives that edited text back.

Why not just a prompt instruction?

The honest caveat: this is a real gate only as long as the approved path is the agent's only path to the side effect — give it the raw credential too and it can route around you.

Self-host or cloud

Cloud — signup above, inbox at app.impri.dev, early beta.

Self-host — full core, MIT, no license key:

git clone https://gitlab.com/sekera.radim/impri.git
cd impri
docker compose up

Server on http://localhost:8484, web inbox on http://localhost:8080. The bootstrap admin key prints to the logs on first start. Details: Self-hosting.

Reference

Everything below is unchanged technical detail: CLI, SDKs, integrations, the full doc set, and legal.

CLI, SDKs & integrations

v0.1, pre-release. MCP is published; the CLI and both SDKs are local-install only (pre-npm / pre-PyPI) for now.

Framework integrations in integrations/: packages for the Claude Agent SDK, CrewAI, LangChain, and the OpenAI Agents SDK; documented webhook-based patterns for n8n, Make, and Zapier in Integrations.

Notifications

Six channels: Slack, Discord, Telegram, email, ntfy, and generic webhook. See Notification channels and Telegram Approval Bot.

Documentation

• Web docs: <https://impri.dev/docs?utm_source=github&utm_medium=readme&utm_campaign=impri>;
• Quickstart — signup to first approved action in under 5 minutes
• Example agent — dependency-free Node script showing the full loop
• Self-hosting — Docker, env vars, backups, reverse proxy
• Webhooks — HMAC verification, retries, polling fallback
• Watcher presets — all 18 templates, REST + SDK + MCP usage
• Audit log — event types, query API, export, retention
• Architecture
• llms.txt — machine-readable index for AI assistants

Pricing

Free (3 watchers, 100 approvals/mo) · Indie $9/mo (20 watchers, 2,000 approvals/mo, 5-minute checks) · Team $29/mo (unlimited watchers and approvals, 1-minute checks). Self-host the full core for free — no tier limits apply outside the hosted cloud. Details: impri.dev.

Privacy & legal

• Privacy Policy
• GDPR notes
• Terms

License

MIT — see LICENSE. Self-host the full core freely; the hosted cloud and paid tiers are the commercial offering (see MONETIZATION.md).

Made by Radim Sekera. Related project: briefgate.dev — client intake for AI coding agents.