ClauVDA - Anthropic Claude AI for NVDA

Summary

ClauVDA integrates Anthropic's Claude AI directly into NVDA, providing blind and visually impaired users with powerful AI assistance. The add-on supports the current Claude line-up — Opus 4.7, Sonnet 4.6, and Haiku 4.5 — for chat, image description, screen-recording analysis, and more. Both the direct Anthropic API and Amazon Bedrock (via bearer-token API keys) are supported as authentication providers.

Features

• AI Chat: Have conversations with Claude directly from NVDA
• Screen Description: Capture and describe the entire screen
• Object Description: Describe the current navigator object
• Video Analysis: Record a short screen clip; Claude analyzes sampled frames
• Attach Images: Attach images from files for AI description
• Conversation History: Maintain context across multiple messages
• Multiple Models: Choose between Opus, Sonnet, and Haiku
• Two Auth Providers: Anthropic API direct, or Amazon Bedrock bearer token
• Summarize Selection: Select text and have Claude summarize the key points
• Customizable Settings: Temperature, max tokens, streaming, and more

Requirements

• NVDA 2024.1 or later
• One of:
• An Anthropic API key, or
• An Amazon Bedrock API key (bearer token) with access to Claude models
• Internet connection

Setup

Option 1 — Anthropic API (direct)

1. Visit console.anthropic.com/settings/keys
2. Create an API key
3. In NVDA: Preferences > Settings > Claude AI
4. Leave "API provider" as Anthropic (direct)
5. Click Configure Anthropic API Key... and paste your key

Option 2 — Amazon Bedrock (bearer token)

1. Ensure your AWS account has model access for the Claude models you plan to use
2. Create a Bedrock API key at the AWS console: Bedrock API keys
3. In NVDA: Preferences > Settings > Claude AI
4. Change "API provider" to Amazon Bedrock
5. Set the AWS region (defaults to us-east-1)
6. Click Configure Bedrock API Key... and paste your bearer token

Keys for each provider are stored separately, encrypted at rest with Windows DPAPI. You can also supply them via environment variable (ANTHROPIC_API_KEY or AWS_BEARER_TOKEN_BEDROCK).

Keyboard Shortcuts

| Shortcut | Action | |----------|--------| | NVDA+G | Open Claude AI dialog | | NVDA+Shift+E | Describe the entire screen | | NVDA+Shift+O | Describe the navigator object | | NVDA+V | Start/stop video recording for analysis | | NVDA+Shift+U | Summarize selected text | | NVDA+Shift+H | Summarize the last spoken text |

Using the Claude Dialog

When you open the Claude dialog with NVDA+G:

1. Model: Select which Claude model to use
2. System Prompt: Optional instructions on how Claude should respond
3. History: View the conversation history
4. Message: Type your message or question
5. Send: Send your message to Claude
6. Attach Image: Add an image file for Claude to analyze
7. Attach Video: Add a video; frames are sampled and sent as images
8. Clear: Clear the conversation history
9. Copy Response: Copy the last response to the clipboard

Dialog Tips

• Press Ctrl+Enter in the message field to quickly send
• Use Tab to navigate between controls
• Alt+1..9,0 reads the Nth most recent message; double-press to copy

Settings

Access settings via NVDA menu > Preferences > Settings > Claude AI:

• API provider: Anthropic direct or Amazon Bedrock
• AWS region: Bedrock region (ignored when using the Anthropic API directly)
• Default Model: Claude model to use by default
• Temperature (0-100): Response randomness (0 = focused, 100 = creative)
• Maximum Output Tokens: Maximum length of responses
• Stream Responses: Display/speak responses as they arrive
• Conversation Mode: Include chat history for context
• Remember System Prompt: Save your custom system prompt
• Block Escape Key: Prevent accidental dialog closure
• Filter Markdown: Remove markdown formatting from responses

Audio Feedback

• Play sound when sending request
• Play sound while waiting
• Play sound when response received

Available Models

• Claude Opus 4.7 — Most capable, extended thinking
• Claude Sonnet 4.6 — Balanced for everyday use, extended thinking
• Claude Haiku 4.5 — Fastest, cost-efficient

All three support image input.

Image and Video Features

Screen Description (NVDA+Shift+E)

Captures your entire screen and sends it to Claude for a detailed description.

Object Description (NVDA+Shift+O)

Captures only the current navigator object.

Video Analysis (NVDA+V)

1. Press NVDA+V to start recording
2. Perform the actions you want to analyze
3. Press NVDA+V again to stop
4. Frames are sampled from the recording and sent to Claude as images

Claude doesn't accept video files directly, so the add-on samples a handful of frames (default: 12) uniformly across the clip.

Summarize Selection (NVDA+Shift+U)

Select text in any application and have Claude summarize the key points.

Troubleshooting

"Anthropic SDK failed to load"

The bundled libraries may be missing or corrupted. Reinstall the add-on.

"No Anthropic API key configured"

Configure your API key in Settings > Claude AI for the provider you've selected.

Responses are too short or cut off

Increase the "Maximum Output Tokens" setting.

Responses are too random

Lower the Temperature setting.

Privacy Notice

• Your messages, images, and video frames are sent to the selected provider (Anthropic or AWS Bedrock)
• API keys are stored locally, encrypted with Windows DPAPI
• No data is shared with the add-on developer
• Review the Anthropic usage policies and/or your AWS Bedrock agreement for details

Building from source

The bundled Python dependencies (Anthropic SDK, boto3, Pillow, mss, imageio, etc.) are not checked into this repository — they are large, platform specific, and easy to reproduce with uv. After cloning you must install them into both addon/lib32 and addon/lib64 before building, otherwise the generated .nvda-addon will not be able to import anthropic at runtime.

```bash

Clone

git clone https://github.com/ogomez92/claudeNVDA.git cd claudeNVDA

Create a local build venv and install build tooling

uv venv uv sync

Install the add-on's runtime dependencies into BOTH lib dirs.

NVDA runs on 64-bit Python by default; 32-bit is kept for older setups.

uv pip install --python 3.13 --target addon/lib64 \ --python-platform x86_64-pc-windows-msvc \ -r addon/requirements.txt

uv pip install --python 3.13 --target addon/lib32 \ --python-platform i686-pc-windows-msvc \ -r addon/requirements.txt

Build the .nvda-addon bundle

uv run scons . ```

The build produces ClauVDA-<version>.nvda-addon in the repo root; install it in NVDA via Tools → Add-on store → Install from external source.

Note: Without running the two uv pip install --target addon/lib* commands, scons will still build an .nvda-addon, but NVDA will log anthropic import failed on load and the add-on will refuse to start.

Support

• Report issues: GitHub Issues
• Source code: GitHub Repository

License

This add-on is released under the GNU General Public License v2.

Author

Oriol Gomez Sentis