turn-mcp-web
Enables human-in-the-loop checkpoints for AI agents, allowing them to pause and wait for human input via a web UI, then seamlessly resume execution.
README
turn-mcp-web
Turn a single agent request into a durable human-in-the-loop conversation.
turn-mcp-web is a self-hosted MCP server with a browser console. An agent calls turn.wait, execution pauses, a human replies in the web UI, and the same run continues without starting over.
Preview
<p align="center"> <img src="./assets/screenshots/live-review-session.png" alt="Live review session" width="100%"> </p>
<p align="center"> <img src="./assets/screenshots/settings-overview.png" alt="Settings overview" width="32%"> <img src="./assets/screenshots/setup-snippets.png" alt="Client setup snippets" width="32%"> <img src="./assets/screenshots/language-picker.png" alt="First launch language picker" width="32%"> </p>
Why It Exists
Most agent integrations treat a human checkpoint as a dead end:
- the model asks a question
- the run stops
- the next message starts a new run with partial context
turn-mcp-web keeps that checkpoint inside the same execution loop.
work -> turn.wait -> human reply -> work -> turn.wait -> human reply
That makes it useful for approvals, branching decisions, operator handoffs, review queues, and long-running assisted workflows.
What You Get
- MCP tool aliases:
turn.wait,turn_wait,turn - browser console for live replies
- Streamable HTTP transport for IDE MCP clients
- stdio entrypoint for desktop clients that spawn MCP servers
- REST long-poll API for Python and non-MCP agent frameworks
- persistent history and event logging
- webhook and Telegram notifications
- operator and viewer API roles
- session list, quick replies, timeout controls, and live SSE updates
- one-command local startup
Quick Start
macOS
Double-click start.command
Windows
Double-click start.bat
Linux
bash start.sh
From Source
npm install
npm run build
npm start
Web console: http://127.0.0.1:3737/
MCP endpoint: http://127.0.0.1:3737/mcp
Core Workflow
- Start the server.
- Connect your MCP client to
http://127.0.0.1:3737/mcpor launch the stdio entrypoint. - Give the agent
skills/turn-mcp-web/SKILL.mdso it routes user-facing checkpoints throughturn.wait. - When the agent pauses, answer in the browser console.
- The same agent run resumes with your reply.
Client Setup
Streamable HTTP
Use this for Cursor, Windsurf, VS Code, Claude Code, Antigravity, and other MCP clients that support remote HTTP servers.
{
"mcpServers": {
"turn-mcp-web": {
"url": "http://127.0.0.1:3737/mcp"
}
}
}
Windsurf uses serverUrl instead of url.
stdio
Use this when the client launches MCP servers as local child processes.
{
"mcpServers": {
"turn-mcp-web": {
"command": "node",
"args": ["/absolute/path/to/dist/server-stdio.js"]
}
}
}
The stdio process still opens the same web console on port 3737.
Python
Use the bundled Python client when your framework does not speak MCP directly.
pip install ./python-client
from turn_mcp_client import TurnMcpClient, TurnMcpCanceled, TurnMcpTimeout
client = TurnMcpClient("http://127.0.0.1:3737")
try:
reply = client.wait(
context="About to apply a production migration.",
question="Should I proceed?",
options=["Proceed", "Show SQL", "Cancel"],
agent_name="MigrationAgent",
)
print(reply)
except TurnMcpTimeout:
print("No reply before timeout.")
except TurnMcpCanceled:
print("Canceled by operator.")
More examples: python-client/README.md
Agent Contract
Give the agent one of these files:
- Skill entry:
skills/turn-mcp-web/SKILL.md - 中文参考:
skills/turn-mcp-web/references/skill.zh-CN.md
Those files tell the agent to use turn.wait as the communication boundary instead of replying directly.
API
Public
GET /healthzGET /api/public-config
Session and wait control
GET /api/waitsGET /api/waits/:idPOST /api/waits/:id/respondPOST /api/waits/:id/cancelPOST /api/waits/:id/extendPOST /api/waits/cancel-allPOST /api/waits/create-and-wait
History and events
GET /api/historyGET /api/history/timelineGET /api/eventsGET /api/stream
Runtime management
GET /api/auth-checkGET /api/sessionsPOST /api/settingsPOST /api/auto-configurePOST /api/auto-unconfigure
Environment
| Variable | Default | Purpose |
|---|---|---|
TURN_MCP_HTTP_HOST |
127.0.0.1 |
HTTP bind host |
TURN_MCP_HTTP_PORT |
3737 |
HTTP bind port |
TURN_MCP_HTTP_PATH |
/mcp |
MCP endpoint path |
TURN_MCP_DEFAULT_TIMEOUT_SECONDS |
600 |
Default wait timeout |
TURN_MCP_API_KEY |
unset | Operator key |
TURN_MCP_VIEWER_API_KEY |
unset | Viewer key |
TURN_MCP_REQUIRE_API_KEY |
auto | Enable auth |
TURN_MCP_EVENT_LOG_FILE |
unset | JSONL event log path |
TURN_MCP_HISTORY_FILE |
unset | JSONL history path |
TURN_MCP_WEBHOOK_URL |
unset | Outbound webhook target |
TURN_MCP_WEBHOOK_EVENTS |
unset | Comma-separated webhook events |
TURN_MCP_WEBHOOK_SECRET |
unset | HMAC signing secret |
TURN_MCP_WEBHOOK_FORMAT |
json |
json, slack, or discord |
TURN_MCP_TELEGRAM_BOT_TOKEN |
unset | Telegram bot token |
TURN_MCP_TELEGRAM_CHAT_ID |
unset | Telegram target chat |
TURN_MCP_TELEGRAM_EVENTS |
wait_created |
Telegram event filter |
TURN_MCP_RATE_LIMIT_MAX |
120 |
Requests per IP |
TURN_MCP_RATE_LIMIT_WINDOW_SECONDS |
60 |
Rate limit window |
TURN_MCP_MAX_CONCURRENT_WAITS_PER_SESSION |
10 |
Per-session concurrency cap |
TURN_MCP_REINFORCEMENT_SUFFIX |
built-in | Appended reminder text |
Authentication
Auth is off by default. When enabled, send either of these headers:
x-turn-mcp-api-key: <key>
Authorization: Bearer <key>
Roles:
operator: full controlviewer: read-only inspection and SSE subscription
Docker
docker build -t turn-mcp-web .
docker run --rm -p 3737:3737 \
-e TURN_MCP_HTTP_HOST=0.0.0.0 \
-e TURN_MCP_API_KEY=your_key \
turn-mcp-web
Compose file: docker-compose.yml
Repository Layout
src/ TypeScript server
public/ Browser console
python-client/ Python client package
assets/ README screenshots and demo video
Contributing
See CONTRIBUTING.md.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.