Qualcomm AI Hub MCP

Compile, profile, and run AI models on real phones and other devices hosted by Qualcomm AI Hub, directly from your MCP client.

Unofficial and in beta. This independent project is not affiliated with, endorsed by, or supported by Qualcomm.

What you can do

Ask your assistant something like this:

Upload model.onnx, compile it for the Samsung Galaxy S24, run it on the device, and tell me how long inference took.

The server covers the Qualcomm AI Hub Core API with 57 tools. You can:

• Browse available devices, chipsets, and frameworks.

• Upload models and datasets from a local workspace.

• Compile, quantize, and link models.

• Run inference on hosted devices and read performance profiles.

• Download compiled models, outputs, and logs.

Here is a real run of that prompt with MobileNetV2, a 14 MB ONNX model:

Device     Samsung Galaxy S24, Android 14
Compile    ONNX to TFLite: SUCCESS
Inference  0.372 ms estimated, on generated input data
Compute    NPU for all 70 layers
Memory     About 172 MB at peak

This measures speed only. It excludes model loading and any preprocessing in an app.

Related MCP server: hedra-mcp-server

Quick start

You need uv and a Qualcomm AI Hub account.

1. Copy your API token from your Qualcomm AI Hub account settings, then save it:

   uvx --from qai-hub==0.57.0 qai-hub configure --api_token YOUR_TOKEN

   This stores the token in ~/.qai_hub/client.ini. On macOS and Linux, restrict that file with chmod 600 ~/.qai_hub/client.ini.

2. Create a folder that holds only the models and data your assistant may use.

3. Add the server to your client. For Claude Code:

   claude mcp add qualcomm -- uvx --from git+https://github.com/oponite/qualcomm-ai-hub-mcp qualcomm-ai-hub-mcp --workspace /absolute/path/to/folder

   See Other clients for VS Code and others.

4. Ask your assistant to list the available Qualcomm devices.

Always set --workspace. Without it, the server uses the folder it started in.

Before you use it

Keep human approval on. Tools that change your account or start jobs require confirm=true, but a model can set that value itself. Require approval in your client for these tools, and do not auto-approve them:

• submit_* and upload_*

• create_project and job_set_name

• *_modify_sharing and *_disable_sharing

Auto-approve modes can skip these prompts, so check your client's permission mode first.

Your models and data leave your machine. Uploads and submissions send them to Qualcomm, where they are stored and processed under your account. Review Qualcomm's terms and pricing first. This server does not track usage or cost.

Your credentials stay out of tool calls. There is no tool for setting a token. The server adds no telemetry of its own, and its remote calls go through the official Qualcomm SDK, which has its own behavior and terms.

Other clients

The server runs locally over stdio. It is not an HTTP endpoint. If your client cannot find uvx, use its absolute path. This project is not on PyPI yet, so these commands install it from GitHub.

VS Code: add this to .vscode/mcp.json:

{
  "servers": {
    "qualcomm": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/oponite/qualcomm-ai-hub-mcp",
        "qualcomm-ai-hub-mcp", "--workspace", "/absolute/path/to/folder"
      ]
    }
  }
}

Clients that use mcpServers:

{
  "mcpServers": {
    "qualcomm": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/oponite/qualcomm-ai-hub-mcp",
        "qualcomm-ai-hub-mcp", "--workspace", "/absolute/path/to/folder"
      ]
    }
  }
}

Server options

Other ways to authenticate

• Set QAIHUB_CLIENT_INI to the path of a configuration file.

• Set QAI_HUB_API_TOKEN in the client's environment. This cannot be combined with --profile.

Limits

• Each release targets one qai-hub version, currently 0.57.0.

• There is no delete tool. Remove uploads, jobs, and projects with Qualcomm's own tools.

• Cancelling a call does not cancel the remote job. After a connection failure, check existing jobs before resubmitting.

• Set a timeout on long waits.

• Models with dynamic input shapes need fixed ones. Pass input_specs when compiling, such as {"input":[1,3,224,224]}.

• Inline data is limited to 16 MiB. Use HDF5 files for larger datasets.

• submit_profile_job and submit_compile_and_profile_jobs are deprecated by the SDK. Use submit_inference_job.

Documentation

• Tool inputs and results

• Security model

• Full tool list

• What has been verified

• Changelog

• Contributing

• Report a security issue

License

Licensed under the Apache License 2.0. This covers this project only. The Qualcomm AI Hub service and qai-hub SDK have their own terms. Qualcomm and Qualcomm AI Hub are trademarks of Qualcomm Incorporated or its subsidiaries.