Env-Doctor

The missing link between your GPU and Python AI libraries

---

"Why does my PyTorch crash with CUDA errors when I just installed it?"

Because your driver supports CUDA 11.8, but pip install torch gave you CUDA 12.4 wheels.

Env-Doctor diagnoses and fixes the #1 frustration in GPU computing: mismatched CUDA versions between your NVIDIA driver, system toolkit, cuDNN, and Python libraries.

It takes 5 seconds to find out if your environment is broken - and exactly how to fix it.

Doctor "Check" (Diagnosis)

Features

Feature	What It Does
One-Command Diagnosis	Check compatibility: GPU Driver → CUDA Toolkit → cuDNN → PyTorch/TensorFlow/JAX
Compute Capability Check	Detect GPU architecture mismatches — catches why torch.cuda.is_available() returns False on new GPUs (e.g. Blackwell) even when driver and CUDA are healthy
Python Version Compatibility	Detect Python version conflicts with AI libraries and dependency cascade impacts
CUDA Auto-Installer	Execute CUDA Toolkit installation directly with --run; CI-friendly with --yes; preview with --dry-run
Safe Install Commands	Get the exact pip install command that works with YOUR driver
Extension Library Support	Install compilation packages (flash-attn, SageAttention, auto-gptq, apex, xformers) with CUDA version matching
AI Model Compatibility	Check if LLMs, Diffusion, or Audio models fit on your GPU before downloading
WSL2 GPU Support	Validate GPU forwarding, detect driver conflicts within WSL2 env for Windows users
Deep CUDA Analysis	Find multiple installations, PATH issues, environment misconfigurations
Container Validation	Catch GPU config errors in Dockerfiles before you build
MCP Server	Expose diagnostics to AI assistants (Claude Desktop, Zed) via Model Context Protocol
CI/CD Ready	JSON output, proper exit codes, and CI-aware env-var persistence (GitHub Actions, GitLab CI, CircleCI, Azure Pipelines, Jenkins)
Fleet Dashboard (optional)	Web UI for monitoring multiple GPU machines — aggregate status, drill-down diagnostics, history timeline. Install with pip install "env-doctor[dashboard]"

Installation

Core CLI

The core CLI has no heavy dependencies — installs in seconds.

pip install env-doctor

# Or with uv (faster, isolated)
uv tool install env-doctor
uvx env-doctor check

With Fleet Dashboard

If you want to manage a distributed system of multiple GPU nodes, this dashboard can help you from a observability POV. It adds a web UI for monitoring multiple GPU machines and has no effect on the core CLI. You will be able to soon take action directly from the dashboard via distributed env-doctor cli instances on each VM!

pip install "env-doctor[dashboard]"

This adds: fastapi, uvicorn, sqlalchemy, aiosqlite

	pip install env-doctor	pip install "env-doctor[dashboard]"
env-doctor check	✅	✅
All CLI commands	✅	✅
MCP server	✅	✅
env-doctor check --report-to	✅	✅
env-doctor report install/status	✅	✅
env-doctor dashboard (web UI)	✗	✅

MCP Server (AI Assistant Integration)

Env-Doctor includes a built-in Model Context Protocol (MCP) server that exposes diagnostic tools to AI assistants like Claude Code and Claude Desktop.

Quick Setup for Claude Desktop

1. Install env-doctor:

   pip install env-doctor
   

2. Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

   {
     "mcpServers": {
       "env-doctor": {
         "command": "env-doctor-mcp"
       }
     }
   }
   

3. Restart Claude Desktop - the tools will be available automatically.

Available Tools (11 Total)

• env_check - Full GPU/CUDA environment diagnostics
• env_check_component - Check specific component (driver, CUDA, cuDNN, etc.)
• python_compat_check - Check Python version compatibility with installed AI libraries
• cuda_info - Detailed CUDA toolkit information
• cudnn_info - Detailed cuDNN library information
• cuda_install - Step-by-step CUDA installation instructions
• install_command - Get safe pip install commands for AI libraries
• model_check - Analyze if AI models fit on your GPU
• model_list - List all available models in database
• dockerfile_validate - Validate Dockerfiles for GPU issues
• docker_compose_validate - Validate docker-compose.yml for GPU configuration

Demo — Claude Code using env-doctor MCP tools

Example Usage

Ask your AI assistant:

• "Check my GPU environment"
• "Is my Python version compatible with my installed AI libraries?"
• "How do I install CUDA Toolkit on Ubuntu?"
• "Get me the pip install command for PyTorch"
• "Can I run Llama 3 70B on my GPU?"
• "Validate this Dockerfile for GPU issues"
• "What CUDA version does my PyTorch require?"
• "Show me detailed CUDA toolkit information"

Learn more: MCP Integration Guide

---

Fleet Dashboard (optional)

The core CLI works standalone. The dashboard is an observability layer for teams running multiple GPU machines.

pip install "env-doctor[dashboard]" unlocks a web UI that aggregates diagnostic results from every machine in your fleet into a single view — no SSH required.

How It Works

There are two roles — the dashboard host (receives and displays reports) and the GPU machines (run checks and send results). They communicate over a simple REST API.

  Dashboard Host                                GPU Machine 1
  ┌─────────────────────-┐                 ┌───────────────────────────────┐
  │ env-doctor dashboard │ ◄──── POST ──── │ env-doctor check --report-to  │
  │ (React UI + SQLite)  │    /api/report  │ (runs locally, POSTs result)  │
  └─────────────────────-┘                 └───────────────────────────────┘
         ▲                                     GPU Machine 2
         │                               ┌───────────────────────────────┐
         └──────────── POST ─────────────│ env-doctor check --report-to  │
                       /api/report       └───────────────────────────────┘

Step 1 — Start the dashboard on any machine with a reachable IP (a cheap CPU instance is enough, no GPU needed):

pip install "env-doctor[dashboard]"
env-doctor dashboard
# → Serving at http://localhost:8765
# → 🔐 Generated new API token at ~/.env-doctor/api-token
#    Token: <copy this — paste into the browser login + share with hosts>

On first launch the dashboard generates a shared API token (saved at ~/.env-doctor/api-token, mode 0600). The browser login screen and every host CLI need that token. Override the location with ENV_DOCTOR_API_TOKEN=<token> in the dashboard's environment.

Step 2 — Report from each GPU machine (only needs the core CLI, not the [dashboard] extra):

pip install env-doctor

# One-time report
env-doctor check --report-to http://<dashboard-host>:8765 --token <token>

# Or: set up automatic reporting every 2 minutes
env-doctor report install \
    --url http://<dashboard-host>:8765 \
    --token <token> \
    --interval 2m

The token is saved in ~/.env-doctor/report-config.json so the scheduled cron / Task Scheduler entry stays clean (no secrets in crontab -l).

report install creates a scheduled task on the GPU machine — a cron job on Linux/macOS or a Windows Task Scheduler entry on Windows. That task runs env-doctor check --report-to <url> on the configured interval.

Smart Change Detection

The scheduled task fires every 2 minutes, but it does not POST every 2 minutes. Each run:

1. Runs env-doctor check locally on the GPU machine
2. Hashes the result (status, checks — excluding timestamps)
3. Compares to the last sent hash stored in ~/.env-doctor/report-state.json

Condition	Action
Hash changed (driver updated, library installed, new issue)	POST full report immediately
Hash unchanged, 30 min since last POST	POST lightweight heartbeat (confirms machine is alive)
Hash unchanged, heartbeat not due	Skip — no network call, sub-second no-op

On a stable machine, this means ~1 POST every 30 minutes instead of 720.

Reporting Commands (run on GPU machines)

These commands run on each GPU machine, not on the dashboard host:

Command	Where	What it does
env-doctor dashboard	Dashboard host	Starts the web UI and API server
env-doctor check --report-to URL	GPU machine	Runs check locally, POSTs result to dashboard
env-doctor report install --url URL	GPU machine	Creates a cron job / scheduled task on this machine
env-doctor report status	GPU machine	Reads this machine's local config and last report time
env-doctor report uninstall	GPU machine	Removes the scheduled task from this machine

report status is purely local — it reads ~/.env-doctor/report-state.json and prints when this machine last reported and whether the scheduler is active. No network call.

Setting Up on Cloud Instances

# AWS / GCP / Azure — on your dashboard VM
pip install "env-doctor[dashboard]"
env-doctor dashboard --host 0.0.0.0 --port 8765
# Open port 8765 in your security group / firewall rules

# On each GPU instance (same VPC) — one command
pip install env-doctor && env-doctor report install --url http://<vm-private-ip>:8765

For machines behind NAT (different networks), use Tailscale for zero-config networking:

# Install Tailscale on each machine, then use the Tailscale IP
env-doctor report install --url http://100.x.x.x:8765 --token <token>

Deploying to a Shared Host

If you're hosting one dashboard for a small team rather than running it on a single laptop:

1. Use the API token. Generated automatically on first launch (see Step 1 above), or pin a known value with ENV_DOCTOR_API_TOKEN. All /api/* routes require Authorization: Bearer <token>.
2. Put TLS in front. The dashboard speaks plain HTTP. Terminate TLS in nginx, Caddy, or a cloud load balancer and forward to 127.0.0.1:8765. Tokens travel in the Authorization header — they need TLS to stay private outside trusted networks.
3. Restrict CORS. Set ENV_DOCTOR_CORS_ORIGINS=https://dashboard.example.com on the dashboard process so the browser only honours requests from your origin (defaults to * for backward compatibility with local use).
4. Tune staleness. A machine is flagged "stale" once its last_seen is older than ENV_DOCTOR_STALE_SECONDS seconds (default 3600 = 1 hour, ≈ 2× the default heartbeat).
5. Rotating the token. Edit ~/.env-doctor/api-token (or update ENV_DOCTOR_API_TOKEN and restart). Each host CLI needs env-doctor report install --token <new> re-run, or its ~/.env-doctor/report-config.json updated, before the next check-in.

What the Dashboard Shows

The web UI at http://<dashboard-host>:8765 displays:

• Topology view — force-directed graph showing the dashboard hub and all GPU machines as nodes, colour-coded by health status, with click-to-zoom detail panels
• Fleet overview — aggregate health pie chart, issues table with expandable per-machine diagnostics, health gauge, and remote command execution
• Machine detail — full diagnostic breakdown identical to what env-doctor check prints locally
• History timeline — past snapshots per machine to track when issues appeared or were resolved

All data stored in ~/.env-doctor/dashboard.db (SQLite) on the dashboard host. No external database or cloud dependencies.

Remote Remediation

The dashboard can queue env-doctor CLI commands to run on remote machines — no SSH required.

  Dashboard                           GPU Machine
  ┌────────────┐                     ┌──────────────────┐
  │ Operator   │                     │ Scheduled check  │
  │ clicks     │                     │ (cron / Task     │
  │ "▶ Run" on │                     │  Scheduler)      │
  │ Fleet page │                     │                  │
  └──────┬─────┘                     └────────┬─────────┘
         │                                     │
         ▼                                     ▼
  Queue command                        env-doctor check
  in database                          --report-to <url>
  (status: pending)                            │
         │                                     ▼
         │                            Server returns pending
         │                            commands in response
         │                                     │
         │                                     ▼
         │                            CLI executes commands,
         │                            posts results back
         │                                     │
         │                                     ▼
         │                            CLI re-runs check to
         └─────────────────────────── verify the fix

How it works: GPU machines check in periodically via env-doctor check --report-to. On each check-in, the server returns any pending commands in the HTTP response. The CLI executes them, posts the output back, and re-runs diagnostics to verify the fix. This pull-based model means:

• No inbound ports needed on GPU machines (works behind NATs, firewalls, VPNs)
• No SSH credentials stored on the dashboard
• Only env-doctor commands are accepted (security scoped)
• Commands typically execute within one check-in interval (default: 5 minutes)

Set up scheduled check-ins on each GPU machine:

# Linux / macOS — creates a cron job
env-doctor report install --url http://<dashboard-host>:8765 --interval 5m

# Windows — creates a Task Scheduler entry
env-doctor report install --url http://<dashboard-host>:8765 --interval 5m

Learn more: Fleet Monitoring Guide

---

Usage

Diagnose Your Environment

env-doctor check

Example output:

🩺 ENV-DOCTOR DIAGNOSIS
============================================================

🖥️  Environment: Native Linux

🎮 GPU Driver
   ✅ NVIDIA Driver: 535.146.02
   └─ Max CUDA: 12.2

🔧 CUDA Toolkit
   ✅ System CUDA: 12.1.1

📦 Python Libraries
   ✅ torch 2.1.0+cu121

✅ All checks passed!

On new-generation GPUs (e.g. RTX 5070 / Blackwell), env-doctor catches architecture mismatches and distinguishes between two failure modes:

Hard failure — torch.cuda.is_available() returns False:

🎯  COMPUTE CAPABILITY CHECK
    GPU: NVIDIA GeForce RTX 5070 (Compute 12.0, Blackwell, sm_120)
    PyTorch compiled for: sm_50, sm_60, sm_70, sm_80, sm_90, compute_90
    ❌ ARCHITECTURE MISMATCH: Your GPU needs sm_120 but PyTorch 2.5.1 doesn't include it.

    This is likely why torch.cuda.is_available() returns False even though
    your driver and CUDA toolkit are working correctly.

    FIX: Install PyTorch nightly with sm_120 support:
       pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu126

Soft failure — torch.cuda.is_available() returns True via NVIDIA's PTX JIT, but complex ops may silently degrade:

🎯  COMPUTE CAPABILITY CHECK
    GPU: NVIDIA GeForce RTX 5070 (Compute 12.0, Blackwell, sm_120)
    PyTorch compiled for: sm_50, sm_60, sm_70, sm_80, sm_90, compute_90
    ⚠️  ARCHITECTURE MISMATCH (Soft): Your GPU needs sm_120 but PyTorch 2.5.1 doesn't include it.

    torch.cuda.is_available() returned True via NVIDIA's driver-level PTX JIT,
    but you may experience degraded performance or failures with complex CUDA ops.

    FIX: Install a newer PyTorch with native sm_120 support for full compatibility:
       pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu126

Check Python Version Compatibility

env-doctor python-compat

🐍  PYTHON VERSION COMPATIBILITY CHECK
============================================================
Python Version: 3.13 (3.13.0)
Libraries Checked: 2

❌  2 compatibility issue(s) found:

    tensorflow:
      tensorflow supports Python <=3.12, but you have Python 3.13
      Note: TensorFlow 2.15+ requires Python 3.9-3.12. Python 3.13 not yet supported.

    torch:
      torch supports Python <=3.12, but you have Python 3.13
      Note: PyTorch 2.x supports Python 3.9-3.12. Python 3.13 support experimental.

⚠️   Dependency Cascades:
    tensorflow [high]: TensorFlow's Python ceiling propagates to keras and tensorboard
      Affected: keras, tensorboard, tensorflow-estimator
    torch [high]: PyTorch's Python version constraint affects all torch ecosystem packages
      Affected: torchvision, torchaudio, triton

💡  Consider using Python 3.12 or lower for full compatibility

💡  Cascade: tensorflow constraint also affects: keras, tensorboard, tensorflow-estimator

💡  Cascade: torch constraint also affects: torchvision, torchaudio, triton

============================================================

Get Safe Install Command

env-doctor install torch

⬇️ Run this command to install the SAFE version:
---------------------------------------------------
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
---------------------------------------------------

Install CUDA Toolkit

Display instructions or execute the installation directly:

# Show platform-specific steps (default)
env-doctor cuda-install

# Preview what would run — no changes made
env-doctor cuda-install --dry-run

# Execute interactively (asks [y/N] before running)
env-doctor cuda-install --run

# Execute headlessly — great for CI/scripts
env-doctor cuda-install --run --yes

# Install a specific version, headless
env-doctor cuda-install 12.6 --run --yes

Example dry-run output (Windows):

[DRY RUN] [1/1] winget install Nvidia.CUDA --version 12.2

[DRY RUN] [1/1] nvcc --version

CUDA 12.2 installation completed successfully.
Verification: PASSED

Full log: C:\Users\you\.env-doctor\install.log

Every run writes a timestamped log to ~/.env-doctor/install.log for debugging.

Supported Platforms:

• Ubuntu 20.04, 22.04, 24.04
• Debian 11, 12
• RHEL 8, 9 / Rocky Linux / AlmaLinux
• Fedora 39+
• WSL2 (Ubuntu)
• Windows 10/11 (via winget)
• Conda (all platforms)

Exit codes for CI pipelines:

Code	Meaning
0	Installation succeeded and verified
1	An installation step failed
2	Installed but nvcc --version failed

Install Compilation Packages (Extension Libraries)

For extension libraries like flash-attn, SageAttention, auto-gptq, apex, and xformers that require compilation from source, env-doctor provides special guidance to handle CUDA version mismatches:

env-doctor install flash-attn

Example output (with CUDA mismatch):

🩺  PRESCRIPTION FOR: flash-attn

⚠️   CUDA VERSION MISMATCH DETECTED
     System nvcc: 12.1.1
     PyTorch CUDA: 12.4.1

🔧  flash-attn requires EXACT CUDA version match for compilation.
    You have TWO options to fix this:

============================================================
📦  OPTION 1: Install PyTorch matching your nvcc (12.1)
============================================================

Trade-offs:
  ✅ No system changes needed
  ✅ Faster to implement
  ❌ Older PyTorch version (may lack new features)

Commands:
  # Uninstall current PyTorch
  pip uninstall torch torchvision torchaudio -y

  # Install PyTorch for CUDA 12.1
  pip install torch --index-url https://download.pytorch.org/whl/cu121

  # Install flash-attn
  pip install flash-attn --no-build-isolation

============================================================
⚙️   OPTION 2: Upgrade nvcc to match PyTorch (12.4)
============================================================

Trade-offs:
  ✅ Keep latest PyTorch
  ✅ Better long-term solution
  ❌ Requires system-level changes
  ❌ Verify driver supports CUDA 12.4

Steps:
  1. Check driver compatibility:
     env-doctor check

  2. Download CUDA Toolkit 12.4:
     https://developer.nvidia.com/cuda-12-4-0-download-archive

  3. Install CUDA Toolkit (follow NVIDIA's platform-specific guide)

  4. Verify installation:
     nvcc --version

  5. Install flash-attn:
     pip install flash-attn --no-build-isolation

============================================================

Check Model Compatibility

env-doctor model llama-3-8b

🤖  Checking: LLAMA-3-8B (8.0B params)

🖥️   Your Hardware: RTX 3090 (24GB)

💾  VRAM Requirements:
  ✅  FP16: 19.2GB - fits with 4.8GB free
  ✅  INT4:  4.8GB - fits with 19.2GB free

✅  This model WILL FIT on your GPU!

List all models: env-doctor model --list

Cloud GPU Recommendations:

# Get cloud GPU recommendations for a model that doesn't fit
env-doctor model llama-3-70b --recommend

# Direct VRAM lookup (no model name needed)
env-doctor model --vram 80000 --recommend

☁️   Cloud GPU Recommendations

  FP16 (~140.0 GB):
    $27.20 /hr  azure  ND96asr_v4              8x A100 (40GB each)          180.0GB free
    $29.39 /hr  gcp    a2-highgpu-8g            8x A100 (40GB each)          180.0GB free
    ...

Automatic HuggingFace Support (New ✨) If a model isn't found locally, env-doctor automatically checks the HuggingFace Hub, fetches its parameter metadata, and caches it locally for future runs — no manual setup required.

# Fetches from HuggingFace on first run, cached afterward
env-doctor model bert-base-uncased
env-doctor model sentence-transformers/all-MiniLM-L6-v2

Output:

🤖  Checking: BERT-BASE-UNCASED
    (Fetched from HuggingFace API - cached for future use)
    Parameters: 0.11B
    HuggingFace: bert-base-uncased

🖥️   Your Hardware:
    RTX 3090 (24GB VRAM)

💾  VRAM Requirements & Compatibility
  ✅  FP16:  264 MB - Fits easily!

💡  Recommendations:
1. Use fp16 for best quality on your GPU

Validate Dockerfiles

env-doctor dockerfile

🐳  DOCKERFILE VALIDATION

❌  Line 1: CPU-only base image: python:3.10
    Fix: FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04

❌  Line 8: PyTorch missing --index-url
    Fix: pip install torch --index-url https://download.pytorch.org/whl/cu121

More Commands

Command	Purpose
env-doctor check	Full environment diagnosis
env-doctor python-compat	Check Python version compatibility with AI libraries
env-doctor cuda-install	Step-by-step CUDA Toolkit installation guide
env-doctor install <lib>	Safe install command for PyTorch/TensorFlow/JAX, extension libraries (flash-attn, auto-gptq, apex, xformers, SageAttention, etc.)
env-doctor model <name>	Check model VRAM requirements
env-doctor cuda-info	Detailed CUDA toolkit analysis
env-doctor cudnn-info	cuDNN library analysis
env-doctor dockerfile	Validate Dockerfile
env-doctor docker-compose	Validate docker-compose.yml
env-doctor init --github-actions	Generate GitHub Actions workflow
env-doctor scan	Scan for deprecated imports
env-doctor debug	Verbose detector output
env-doctor dashboard	Start fleet monitoring web UI (requires [dashboard] extra)
env-doctor report install	Set up periodic reporting via cron (Linux) or Task Scheduler (Windows)

CI/CD Integration

Generate a GitHub Actions workflow with one command:

env-doctor init --github-actions

This creates .github/workflows/env-doctor.yml — review, commit, and push. Your CI will validate the ML environment on every push and PR.

Or add manually:

# JSON output for scripting
env-doctor check --json

# CI mode with exit codes (0=pass, 1=warn, 2=error)
env-doctor check --ci

Exit code semantics for check --json / check --ci:

Code	Meaning
0	All detected components are compatible (uninstalled libraries do not count as failures)
1	Installed components have warnings or version conflicts
2	One or more components are in an error state

GitHub Actions example:

- run: pip install env-doctor
- run: env-doctor check --ci

Documentation

Full documentation: https://mitulgarg.github.io/env-doctor/

• Getting Started
• Command Reference
• MCP Integration Guide
• WSL2 GPU Guide
• CI/CD Integration
• Architecture

Video Tutorial: Watch Demo on YouTube

Platform Support

Env-Doctor is built for Linux and Windows — the platforms where NVIDIA GPUs and CUDA are available. All GPU diagnostics (driver, CUDA, cuDNN, library compatibility) target these platforms.

macOS is supported for non-GPU features: Fleet Dashboard hosting, model memory checks (env-doctor model), Python compatibility, project import scanning, and the MCP server. This makes a Mac a great centralised dashboard host while your Linux/Windows VMs handle the GPU workloads.

macOS uses zsh, which treats [] as a glob pattern. Quote extras when installing: pip install "env-doctor[dashboard]"

Contributing

Contributions welcome! See CONTRIBUTING.md for details.

License

MIT License - see LICENSE