ask-human mcp 🧑‍💻🤝🤖

stop your ai from hallucinating. gives it an escape route when confused instead of false confidence.

the pain

ai blurts out an endpoint that never existed

the agent makes assumptions that are simply not true and has false confidence

repeat x100 errors and your day is spent debugging false confidence and issues when you could simply ask a question

the fix

an mcp server that lets the agent raise its hand instead of hallucinating. feels like mentoring a sharp intern who actually asks before guessing.

agent → ask_human()

⬇

question lands in ask_human.md

⬇

you swap "PENDING" for the answer

⬇

agent keeps coding

sample file:

### Q8c4f1e2a
ts: 2025-01-15 14:30  
q: which auth endpoint do we use?  
ctx: building login form in auth.js  
answer: PENDING

you drop:

answer: POST /api/v2/auth/login

boom. flow continues and hopefully the issues are solved.

why it's good

• pip install ask-human-mcp → done
• zero config, cross-platform
• watches the file, instant feedback
• multiple agents, no sweat
• locks + limits so nothing catches fire
• full q&a history in markdown (nice paper-trail for debugging)

30-sec setup

pip install ask-human-mcp
ask-human-mcp

.cursor/mcp.json:

{
  "mcpServers": {
    "ask-human": { "command": "ask-human-mcp" }
  }
}

restart cursor and vibe.

how it works

1. ai gets stuck → calls ask_human(question, context)

2. question logged → appears in ask_human.md with unique ID

3. human answers → replace "PENDING" with your response

4. ai continues → uses your answer to proceed

the ai receives your answer and keeps coding!

config options (if you want them)

command line

ask-human-mcp --help
ask-human-mcp --port 3000 --host 0.0.0.0  # http mode
ask-human-mcp --timeout 1800               # 30min timeout  
ask-human-mcp --file custom_qa.md          # custom q&a file
ask-human-mcp --max-pending 50             # max concurrent questions
ask-human-mcp --max-question-length 5000   # max question size
ask-human-mcp --rotation-size 10485760     # rotate file at 10mb

different clients

cursor (local):

{
  "mcpServers": {
    "ask-human": {
      "command": "ask-human-mcp",
      "args": ["--timeout", "900"]
    }
  }
}

cursor (http):

{
  "mcpServers": {
    "ask-human": {
      "url": "http://localhost:3000/sse"
    }
  }
}

claude desktop:

{
  "mcpServers": {
    "ask-human": {
      "command": "ask-human-mcp"
    }
  }
}

what's in the box

• zero configuration → works out of the box
• file watching → instant response when you save answers
• timeout handling → questions don't hang forever
• concurrent questions → handle multiple ai agents
• persistent logging → full q&a history in markdown
• cross-platform → windows, macos, linux
• mcp standard → works with any mcp client
• input validation → size limits and sanitization
• file rotation → automatic archiving of large files
• resource limits → prevent dos and memory leaks
• robust parsing → handles malformed markdown gracefully

security stuff

• input sanitization → removes control characters and validates sizes
• file locking → prevents corruption from concurrent access
• secure permissions → files created with restricted access
• resource limits → prevents memory exhaustion and dos attacks
• path validation → ensures files are written to safe locations

limits (so nothing breaks)

platform support

• windows → full support with native file locking
• macos → full support with fsevents file watching
• linux → full support with inotify file watching

api stuff

ask_human(question, context="")

ask the human a question and wait for response.

answer = await ask_human(
    "what database should i use for this project?",
    "building a chat app with 1000+ concurrent users"
)

other tools

• list_pending_questions() → get questions waiting for answers
• get_qa_stats() → get stats about the q&a session

development

from source

git clone https://github.com/masonyarbrough/ask-human-mcp.git
cd ask-human-mcp
pip install -e ".[dev]"
ask-human-mcp

tests

pytest tests/ -v

code quality

black ask_human_mcp tests
ruff check ask_human_mcp tests  
mypy ask_human_mcp

contributing

would love any contributors

issues

use the github issue tracker to report bugs or request features.

you can also just email me: mason@kallro.com

include:

• python version
• operating system
• mcp client (cursor, claude desktop, etc.)
• error messages or logs
• steps to reproduce

changelog

see CHANGELOG.md for version history.

license

mit license - see LICENSE file for details.

thanks

• model context protocol for the excellent standard
• anthropic for claude and mcp support
• cursor for mcp integration
• all contributors and users providing feedback