InstrMCP: Instrumentation Control MCP Server

MCP server suite for quantum device physics laboratory's instrumentation control, enabling Large Language Models to interact directly with physics instruments and measurement systems through QCodes and JupyterLab.

https://github.com/user-attachments/assets/e7d0a441-36b2-4fec-9c54-1427310b7698

Features

• Full QCodes Integration: Built-in support for all QCodes instrument drivers

• Database Integration: Read-only access to QCodes databases with intelligent code generation

• MeasureIt Templates: Comprehensive measurement pattern library and code generation

• JupyterLab Native: Seamless integration with JupyterLab

• Kernel Awareness: notebook_kernel_status / notebook_wait_for_kernel report whether the kernel is busy and wait for it to go idle - working even while a cell is stalled

• Dynamic Tool Creation: Create custom MCP tools at runtime using LLM-powered tool registration

• Safe mode: Read-only mode with optional unsafe execution

• GUI control panel: instrmcp app opens a Streamlit dashboard to launch JupyterLab and watch live status, logs, and MeasureIt sweeps

• Embedded MCP Inspector: the GUI's Inspector tab browses and calls the server's tools/resources/prompts — a native, Node-free alternative to the official npx inspector

• CLI: Easy server management with instrmcp command

• MCP: Standard Model Context Protocol for LLM integration

• The MCP has been tested to work with Claude Desktop, Claude Code, and Codex CLI

Related MCP server: axiomatic-mcp

Quick Start

Installation

From PyPI (Recommended):

pip install instrmcp

That's it! QCodes, JupyterLab, and all dependencies are automatically installed. The JupyterLab extension is automatically enabled (no Node.js or rebuild required).

From Source (For Development):

git clone https://github.com/caidish/instrMCP.git
cd instrMCP
pip install -e .

# Run setup to enable JupyterLab extension (only needed for editable install)
instrmcp-setup

# Set required environment variable for development
# For macOS/Linux:
export instrMCP_PATH="$(pwd)"
echo 'export instrMCP_PATH="'$(pwd)'"' >> ~/.zshrc  # or ~/.bashrc
source ~/.zshrc

# For Windows (PowerShell):
$env:instrMCP_PATH = (Get-Location).Path
[System.Environment]::SetEnvironmentVariable('instrMCP_PATH', (Get-Location).Path, 'User')

Extension: MeasureIt

MeasureIt provides comprehensive measurement pattern templates for common physics experiments.

Installation:

pip install qmeasure

Important Notes:

• Import as measureit (not qmeasure): import measureit

• Python 3.8+ required

• For source installation or advanced configuration, see the MeasureIt GitHub repository

Enable in InstrMCP:

# In Jupyter notebook
%mcp_option add measureit
%mcp_restart

Usage

Loading InstrMCP in Jupyter

# Start JupyterLab
jupyter lab

In a Jupyter notebook cell:

# Load InstrMCP extension
%load_ext instrmcp.extensions

# Start MCP server
%mcp_start

# Check status
%mcp_status

# Enable unsafe mode (code execution)
%mcp_unsafe

# Enable optional features (restart required)
%mcp_option add measureit database
%mcp_restart

You don't need to type the magics. The JupyterLab extension auto-loads the InstrMCP kernel extension in every notebook, and the notebook toolbar has a one-click Start button (safe mode by default). The cell commands above are just the manual equivalent.

Friendly GUI (Streamlit control panel)

For a no-terminal experience, use the Streamlit control panel:

pip install 'instrmcp[gui]'     # one-time: install the GUI extra
instrmcp app --profile demo     # opens http://localhost:8501

From the GUI you can launch JupyterLab, watch live status / logs / MeasureIt, run diagnostics, and recover (restart kernel / stop) — all in one pane. To start the MCP server, open a notebook from the GUI's Open JupyterLab link and click Start in the InstrMCP toolbar (safe mode). The GUI then shows MCP → ready.

The GUI also has an 🔍 Inspector tab — a built-in, Node-free MCP Inspector. Once MCP is ready, click Connect / Refresh to browse the server's tools, resources, and prompts, fill in a JSON-argument form, Call a tool (or read a resource / render a prompt), and view the raw result. It talks the same MCP protocol as the kernel-hosted server (127.0.0.1:8123/mcp), so it needs no npx, no Node.js, and no extra ports — unlike the official MCP Inspector.

Command-line launcher (optional)

The same supervisor is available headlessly, with a built-in HTML dashboard at http://127.0.0.1:8124/:

instrmcp doctor --profile demo      # readable environment diagnostics + fixes
instrmcp launch --profile demo      # launch JupyterLab + supervisor (foreground)
instrmcp status --profile demo      # component states (or --json)
instrmcp logs --follow              # stream JupyterLab / supervisor logs
instrmcp restart --component kernel # restart kernels via the Jupyter REST API
instrmcp stop                       # shut down JupyterLab + supervisor
instrmcp profiles list              # discover bundled / user / project profiles

The supervisor observes runtime health (JupyterLab, MCP reachability) — it never owns the MCP lifecycle. Start/stop/mode for MCP stay in the JupyterLab toolbar.

Profiles are YAML, deep-merged over a bundled default. Search order: project-local (./.instrmcp/profiles/<name>.yaml) → user (~/.instrmcp/profiles/<name>.yaml) → bundled. See instrmcp profiles show <name>.

Optionally register an instrmcp kernelspec whose kernel auto-starts the MCP server on launch (no toolbar click). This is not required — the toolbar works with any kernel — and adds a kernel-selection step:

instrmcp install-kernel --profile demo   # registers the "instrmcp" kernel
instrmcp uninstall-kernel                 # remove it

Then open notebooks on the "Python 3 (instrmcp · …)" kernel.

CLI Utilities

instrmcp config            # Show configuration paths
instrmcp version           # Show version
instrmcp metadata tokens   # Count tokens in metadata descriptions
instrmcp --help            # Show all commands

Documentation

• Architecture - Technical architecture, package structure, MCP tools and resources

• Troubleshooting - Common issues and solutions

• Development Guide - Development setup, testing, code quality, contributing

  • Includes Threading Architecture & Qt Integration - How IPython kernel, Qt event loop, and MCP server thread interact; what cross-thread communication approaches work and don't work with MeasureIt

Configuration

View current configuration:

instrmcp config

Environment variables

• INSTRMCP_FRONTEND_TIMEOUT — seconds the backend waits for a JupyterLab frontend response on cell operations (add/move/read that round-trip through the bridge). Default 10. Raise it on slow machines or heavy notebooks; adding a markdown cell in particular needs an extra frontend round-trip, so a value that is too low can report a false "Timeout waiting for frontend response" even though the cell was added. Read once at kernel start (restart to apply).

Claude Desktop Integration

InstrMCP provides seamless integration with Claude Desktop for AI-assisted laboratory instrumentation control.

Quick Setup (2 Steps)

1. Run Automated Setup:

cd /path/to/your/instrMCP
./agentsetting/claudedesktopsetting/setup_claude.sh

2. Restart Claude Desktop completely and test with: "What MCP tools are available?"

Manual Setup Alternative:

# 1. Copy and edit configuration
cp agentsetting/claudedesktopsetting/claude_desktop_config.json ~/Library/Application\ Support/Claude/claude_desktop_config.json

# 2. Edit the copied file - replace placeholders with actual paths:
#    /path/to/your/python3 → $(which python3)
#    /path/to/your/instrMCP → $(pwd)

See agentsetting/claudedesktopsetting/README.md for detailed setup instructions and troubleshooting.

Claude Code Integration

Claude Code supports local MCP servers via STDIO. Use the provided launcher to connect:

# Add instrMCP as MCP Server
claude mcp add instrMCP --env instrMCP_PATH=$instrMCP_PATH \
  --env PYTHONPATH=$instrMCP_PATH \
  -- $instrMCP_PATH/venv/bin/python \
  $instrMCP_PATH/agentsetting/claudedesktopsetting/claude_launcher.py

# Verify connection
/mcp

Prerequisites:

• Ensure instrMCP_PATH environment variable is set

• Have a Jupyter server running with the instrMCP extension loaded

• MCP server started in Jupyter with %mcp_start

Codex CLI Integration

Codex expects MCP servers over STDIO. Use the Codex launcher to proxy STDIO calls to your HTTP MCP server.

Configuration:

• command: python

• args: ["/path/to/your/instrMCP/agentsetting/codexsetting/codex_launcher.py"]

• env:

  • JUPYTER_MCP_HOST=127.0.0.1

  • JUPYTER_MCP_PORT=8123

Gemini CLI Integration

Gemini CLI supports MCP servers over STDIO. Use the same launcher as Claude Desktop:

Configuration (~/.gemini/settings.json):

{
  "mcpServers": {
    "instrMCP": {
      "command": "/path/to/your/python",
      "args": ["/path/to/your/instrMCP/agentsetting/claudedesktopsetting/claude_launcher.py"],
      "env": {
        "instrMCP_PATH": "/path/to/your/instrMCP",
        "PYTHONPATH": "/path/to/your/instrMCP"
      },
      "trust": true
    }
  }
}

See agentsetting/geminisetting/README.md for detailed setup instructions.

V2.0.0 Features (Current Release)

1. Resource Discovery Tool

The mcp_list_resources() tool helps LLMs discover and effectively use MCP resources:

Features:

• Comprehensive Resource Listing: All available MCP resources with URIs, descriptions, and use cases

• Context-Aware: Only shows resources based on enabled options (core, MeasureIt, database)

• Resources vs Tools Guidance: Educates LLMs on when to use read-only resources vs active tools

• Common Patterns: Examples like "Check available_instruments → Use qcodes_instrument_info"

• First-Use Recommendation: Use this tool FIRST to discover what context is available

Example Response:

{
  "total_resources": 8,
  "resources": [
    {
      "uri": "resource://available_instruments",
      "name": "Available Instruments",
      "use_when": "Need to know what instruments exist BEFORE calling qcodes_instrument_info",
      "example": "Check this first to see instrument names..."
    }
  ],
  "guidance": {
    "resources_vs_tools": {
      "resources": "Provide READ-ONLY reference data, templates, and documentation",
      "tools": "Perform ACTIVE operations like reading live values, executing code"
    },
    "when_to_use_resources": [
      "Before using tools - check available_instruments first",
      "For code templates - get MeasureIt examples",
      "For configuration - check database_config"
    ]
  }
}

2. Consent System for Cell Modifications

Cell modification tools now require user consent in unsafe mode:

Tools requiring consent:

• notebook_update_editing_cell - Shows old/new content comparison before replacing entire cell

• notebook_apply_patch - Shows visual diff dialog with exact changes

• notebook_execute_code - Runs a passed code string directly on the kernel, bypassing the JupyterLab frontend bridge (no cell added). Recovery path when the notebook bridge is degraded but the kernel is alive.

Features:

• Visual Diff Display: Red deletions, green additions, context lines

• Pattern Warning: Prominent alert if old_text not found in cell

• Change Statistics: Shows chars removed/added and delta

• Consent Workflow: "Decline" | "Allow" | "Always Allow" buttons

• Battle-Tested Diffing: Uses industry-standard diff library (v8.0.2) from GitHub/GitLab

Example: When LLM calls notebook_apply_patch(old_text="x = 10", new_text="x = 20"), user sees:

- x = 10  ← Red background with strikethrough
+ x = 20  ← Green background

3. Line Range Parameters for Context Management

Control LLM context window consumption with line range selection:

Features:

• line_start / line_end parameters (default: lines 1-100)

• Automatic Bounds Clamping: Invalid ranges safely handled

• Truncation Metadata: Returns total_lines, truncated flag

• Context Window Savings: Prevents large cells from consuming excessive tokens

Example:

# Get only first 50 lines of a large cell
get_editing_cell(line_start=1, line_end=50)

# Get lines 100-200 for focused analysis
get_editing_cell(line_start=100, line_end=200)

4. Dynamic Tool Creation

Create custom MCP tools at runtime using LLM-powered tool registration:

# In Jupyter with instrMCP loaded in unsafe mode
# LLM can create tools dynamically using meta-tools:
dynamic_register_tool(
    name="analyze_data",
    source_code="def analyze_data(x): return x * 2",
    capabilities=["cap:numpy", "cap:custom.analysis"],  # Freeform labels
    parameters=[{"name": "x", "type": "number", "description": "Input", "required": true}]
)

Features:

• 6 Meta-Tools: register, update, revoke, list, inspect, registry_stats

• Consent UI: User approval required for tool registration/updates (shows full source code)

• Freeform Capability Labels: Tag tools with descriptive capabilities for discovery

• Persistent Registry: Tools saved to ~/.instrmcp/registry/ and reloaded on server start

• Audit Trail: All tool operations logged to ~/.instrmcp/audit/tool_audit.log

• Auto JSON Correction: Optional LLM-powered JSON error fixing (opt-in via %mcp_option auto_correct_json)

Capability Labels (v2.0.0): Capabilities are freeform documentation labels - NOT enforced security boundaries. Use any descriptive string:

• Suggested format: cap:library.action (e.g., cap:numpy.array, cap:qcodes.read)

• Used for discovery, filtering, and transparency in consent UI

• No validation - flexibility for LLMs to describe tool dependencies

• Future: Enforcement layer planned for v3.0.0

See Dynamic Tools Quickstart for details.

Testing & Quality

• Unit tests: Comprehensive coverage of core functionality

• E2E tests: 166 Playwright tests (164 passed, 2 skipped) covering:

  • Server lifecycle and mode switching

  • Safe/unsafe/dangerous mode tools

  • Security scanner pattern blocking

  • Optional features (MeasureIt, Database, Dynamic Tools)

  • Frontend widget and consent dialogs

• Zero linter errors on critical checks

• Code formatted with black

• CI/CD passing on all workflows

See tests/e2e/README.md for E2E test documentation.

V3.0.0 Roadmap

• Capability Enforcement: Security boundaries based on capability taxonomy

• Support RedPitaya

• Support Raspberry Pi for outdated instruments

• Integrating lab wiki knowledge base for safety rails

• More LLM integration examples

License

MIT License - see LICENSE file.

Contributing

We welcome contributions! See our Development Guide for details on:

• Setting up development environment

• Running tests

• Code quality standards

• Contribution guidelines

Links

• Documentation

• Issues

• QCodes

• Model Context Protocol