Your AI agent is about to delete the production database. Do you want it to just go ahead?
call-a-human-mcp is an MCP server that gives any AI agent a pause button — it can ask you a question or request your approval before taking action, and it won't proceed until you respond.
Claude: request_approval("Drop table users_backup — 2.1GB, irreversible")
Slack: ⚠️ AI Agent requesting approval
Action: Drop table users_backup — 2.1GB, irreversible
[Approve] [Deny]
← you click Deny
Claude: "Understood, skipping the deletion."
MCP server that gives any AI phone a friend options : ask a human or request approval via Slack or macOS dialogs before taking irreversible actions.
2026-10-27
ดาว
15
เครื่องมือ
—
อัปไทม์
—
Works with Claude Desktop, Cursor, Windsurf, and any MCP-compatible agent. Notifications via Slack, Telegram, or macOS system dialogs.
Two tools:
Tool
When to use
Returns
ask_human(question, context?)
Need information only a human can provide
str — human's text reply
request_approval(action, details?)
Before any irreversible action
{"approved": bool, "reason": str}
The tool call blocks until you respond (or the timeout expires).
Use the full path to uv, not just uv. Claude Desktop launches with a restricted PATH that won't find uv in ~/.local/bin. Run which uv to get your full path.
4. Restart Claude Desktop (quit fully — Cmd+Q — then reopen).
What you'll see:
When Claude calls ask_human, a native macOS dialog appears:
┌─────────────────────────────────────────────────────┐
│ Claude is asking: │
│ Which database environment should I target? │
│ │
│ [ Reply here... ] │
│ [Cancel] [OK] │
└─────────────────────────────────────────────────────┘
When Claude calls request_approval, a dialog with Approve/Deny options appears. Claude blocks until you respond.
On Linux/Windows or CI? No interactive fallback exists without a terminal. Use Telegram or Slack instead.
Option B: Telegram
⚠️ Not extensively tested. The implementation follows the Telegram Bot API spec and basic flows work, but edge cases may exist. Feedback welcome.
Best for personal use — instant phone notifications, buttons work in the Telegram app.
The number is your chat ID (negative for groups, e.g. -100123456789).
3. Verify credentials:
CALL_HUMAN_CHANNEL=telegram \
TELEGRAM_BOT_TOKEN=<token> \
TELEGRAM_CHAT_ID=<chat_id> \
uv run call-a-human-mcp --check
Expected output:
Checking telegram channel...
Bot token: OK (bot: @your_bot_username)
Test message: OK (chat_id: -100123456789)
Telegram check passed. call-a-human-mcp is ready to use.
A test message also appears in your Telegram chat. If it doesn't, recheck the token and chat ID.
When Claude calls ask_human, a message appears in your Telegram chat:
🤔 Claude is asking:
Which database environment should I target?
Context: Running migration job started at 14:32.
Reply to this message with your answer.
Reply directly to the message. Claude receives your reply and continues.
When Claude calls request_approval, you get Approve/Deny buttons:
⚠️ Approval requested:
Deploy api-service v2.4.1 to production
Details: Replaces v2.3.8. 12 pods will restart.
[ ✅ Approve ] [ ❌ Deny ]
Tap a button — Claude immediately receives the result.
Option C: Slack
Best for teams — Approve/Deny buttons, messages stay in your team's channel.
Security note: The SSE transport has no built-in authentication. Protect it with a reverse proxy (nginx, Caddy) or firewall rules — anyone who can reach the port can send messages to your Slack/Telegram channel.
For self-hosted deployments or clients that connect over HTTP:
cp .env.example .env # fill in your credentials
docker compose up -d
Audit logs are written to ./logs/audit.jsonl on the host.
Getting Claude to call these tools automatically
The MCP server already tells Claude when to use the tools, but the most reliable way to make Claude call them proactively — without you explicitly asking — is to add a custom system prompt in your AI client.
The key principle (learned from production use): tell Claude to call the tools directly — not to ask you whether to call them. The approval happens in Slack/Telegram. Claude's job is just to trigger it.
Claude Desktop
Go to Settings → Custom Instructions and add:
You have access to request_approval and ask_human tools via the call-a-human MCP server.
Call request_approval BEFORE any irreversible action: deleting files, sending
messages, making purchases, modifying production systems, running destructive
commands. Do NOT ask "should I proceed?" — just call the tool and wait.
Only continue if you receive {"approved": true}.
Call ask_human when you are unsure about preferences, file paths, credentials,
or any ambiguous decision. Never guess — ask.
This makes the behavior consistent across all conversations, without needing to remind Claude each time.
Cursor / Windsurf
Add a .cursorrules file (Cursor) or equivalent to your project:
Before any irreversible action, call the request_approval MCP tool directly —
do not ask the user whether to call it. Wait for {"approved": true} before proceeding.
When unsure about preferences or credentials, call ask_human instead of guessing.
Trying tools interactively (without an AI agent)
Use the MCP Inspector to call tools directly:
CALL_HUMAN_CHANNEL=cli uv run mcp dev src/call_a_human_mcp/server.py
The browser UI lets you call ask_human and request_approval manually and inspect the responses.
All configuration options
Variable
Required
Default
Description
CALL_HUMAN_CHANNEL
Yes
—
cli, slack, or telegram
CALL_HUMAN_TIMEOUT
No
300
Seconds to wait before auto-denying
CALL_HUMAN_AUDIT_LOG
No
—
Path to JSONL audit log file
SLACK_BOT_TOKEN
Slack only
—
Bot OAuth token (xoxb-…)
SLACK_APP_TOKEN
Slack only
—
Socket Mode app token (xapp-…)
SLACK_CHANNEL_ID
Slack only
—
Channel to post into (C…)
TELEGRAM_BOT_TOKEN
Telegram only
—
Bot token from @BotFather
TELEGRAM_CHAT_ID
Telegram only
—
Chat/group ID to post into
Copy .env.example to .env and fill in your values.
Audit log
Set CALL_HUMAN_AUDIT_LOG to enable append-only JSONL logging:
AI agent (Claude) call-a-human-mcp Human (Slack/Telegram/macOS)
───────────────── ──────────────── ────────────────────────────
request_approval( block on sees message with
"delete database") ──► threading.Event ──► Approve / Deny buttons
│
│ clicks Approve
▼
{"approved": true, ◄── event.set() ◄── button/dialog handler fires
"reason": "alice"}
The MCP tool handler blocks on a threading.Event. A background daemon thread (Slack Socket Mode, Telegram long-poll, or macOS dialog subprocess) fires event.set() when the human responds.
Development
git clone https://github.com/nishantmodak/call-a-human-mcp
cd call-a-human-mcp
uv sync --extra dev
uv run --extra dev pytest -v
uv run --extra dev ruff check src tests
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
Call A Human ต้องใช้ API key ไหม?
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Call A Human ใช้ทรานสปอร์ตแบบใด?
Call A Human ทำงานแบบโลคัลผ่าน stdio
Call A Human ถูกตรวจสอบล่าสุดเมื่อใด?
ตรวจสอบล่าสุดเมื่อ 2026-10-27 MCP server that gives any AI phone a friend options : ask a human or request approval via Slack or macOS dialogs before taking irreversible actions.