> ## Documentation Index
> Fetch the complete documentation index at: https://dragonwingdocs.qualcomm.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Integrated MCP Server

> Use the built-in Model Context Protocol server to let AI assistants directly control tools for device management, AI models, projects, and more.

The Qualcomm VS Code Extension ships with an **embedded MCP (Model Context Protocol) server** that starts automatically when the extension loads. AI coding assistants such as Claude, Codex, Cursor, GitHub Copilot, etc. can connect to it over `localhost` and invoke a structured tool call — no manual configuration required.

![MCP Server running indicator in the extension sidebar](https://mintcdn.com/qualcomm-prod/hlNuX08-pFFtVn68/archives/2.1/Tools/QVSCE/images/mcp_server_running.png?fit=max\&auto=format\&n=hlNuX08-pFFtVn68\&q=85\&s=966c9b0389c0b34ddbc636ae826ddf3b)

> [!NOTE]
>
> The MCP server runs entirely on your local machine. No data leaves your device
> through the MCP channel; all tool calls go to `127.0.0.1` only.

***

## How It Works

When enabled, the extension starts an HTTP server on `localhost`.

```
http://127.0.0.1:3100/mcp
```

![MCP Server: Running status shown in the extension sidebar navigation](https://mintcdn.com/qualcomm-prod/hlNuX08-pFFtVn68/archives/2.1/Tools/QVSCE/images/mcp_server_status.png?fit=max\&auto=format\&n=hlNuX08-pFFtVn68\&q=85\&s=470cfc626c85e622559cd88c948cea59)

***

## Configuration

The sidebar shows **MCP Server: Running (port 3100)** when active. Click this label to open the MCP & C++ RAG settings directly.

Alternatively, Open VS Code Settings and search for `@id:qvsce.mcpServer.enabled @id:qvsce.mcpServer.port @id:qvsce.cppRag.samplesDir` to see all three MCP-related settings in one view.

![VS Code Settings panel showing QVSCE MCP server configuration options](https://mintcdn.com/qualcomm-prod/hlNuX08-pFFtVn68/archives/2.1/Tools/QVSCE/images/mcp_server_settings.png?fit=max\&auto=format\&n=hlNuX08-pFFtVn68\&q=85\&s=59476b02f16ba5f0122689528e201121)

| Setting | Default | Description |
| - | - | - |
| `qvsce.mcpServer.enabled` | `true` | Enable or disable the embedded MCP server. Changes take effect immediately — no window reload needed. |
| `qvsce.mcpServer.port` | `3100` | Port the HTTP server listens on (1–65535). Change this if port 3100 conflicts with another process. |
| `qvsce.cppRag.samplesDir` | *(empty)* | Absolute path to a directory of C/C++ source files (`.cpp`, `.cc`, `.cxx`, `.h`, `.hpp`, `.hxx`) representing your preferred coding style, SDK-specific patterns, or domain-specific application code — for example, internal SDK wrapper libraries, team utility code, or reference applications. All matching files in subdirectories are discovered recursively. After setting this path, call `build_cpp_rag_index` to populate the index. Leave empty to rely on the built-in default index only. |

> [!TIP]
>
> After changing the port, any MCP client configuration pointing to the old port
> (e.g., a `claude_desktop_config.json` entry) must be updated to match.

### Connecting an AI Client

Any MCP-compatible AI assistant can connect to the server. The general pattern is the same across clients: point it at `http://127.0.0.1:3100/mcp` using the **Streamable HTTP** transport.

**1. Confirm the server is running**

Check the extension sidebar for **MCP Server: Running (port 3100)**. If it shows Stopped, enable it in Settings (`qvsce.mcpServer.enabled`).

**2. Configure your AI client**

Add the server endpoint to your client's MCP configuration. See the examples below for Claude and Codex — other clients follow the same pattern.

> [tab] Claude Desktop / Claude Code
>
> Add to `claude_desktop_config.json` (Claude Desktop) or `.claude/mcp.json` (Claude Code in a project):
>
> ```json theme={null}
> {
>   "mcpServers": {
>     "qualcomm-ide": {
>       "url": "http://127.0.0.1:3100/mcp"
>     }
>   }
> }
> ```
>
> In **Claude Code** you can also add it from the terminal:
>
> ```bash theme={null}
> claude mcp add qualcomm-ide --transport http http://127.0.0.1:3100/mcp
> ```

> [tab] Codex CLI
>
> Add to `~/.codex/config.toml` or pass inline:
>
> ```toml theme={null}
> [[mcp_servers]]
> name = "qualcomm-ide"
> url  = "http://127.0.0.1:3100/mcp"
> ```
>
> Or pass it as a flag when running Codex:
>
> ```bash theme={null}
> codex --mcp-server "qualcomm-ide=http://127.0.0.1:3100/mcp"
> ```

> [tab] Other clients
>
> Any client that supports the MCP Streamable HTTP transport can connect. Use:
>
> * **URL:** `http://127.0.0.1:3100/mcp`
> * **Transport:** Streamable HTTP (session-based)
> * **Auth:** None — local-only, no credentials required
>
> Consult your client's documentation for where to add MCP server entries.

**3. Verify the connection**

Ask your assistant: *"List my connected Qualcomm devices."* A response with device information confirms the server is reachable and tools are available.

***

## Available Tools

The server exposes tools across nine functional categories. The tool set continues to evolve — to see the latest available tools, ask your AI assistant: *"What MCP tools are available from the Qualcomm IDE?"* Click a category below to jump to its tool reference. Tools marked **Linux-only** are unavailable on Windows-native VS Code (WSL remote connections are supported).

> [card] [Device Management](#device-management)
>
> *"What devices are connected right now?"*
>
> *"Connect to my device over SSH at 192.168.1.42 as root."*
>
> *"Flash the latest Qualcomm Linux image onto my active device."*
>
> *"Scan for Wi-Fi networks on my device and connect it to my network."*

> [card] [Projects](#projects)
>
> *"Build my project and deploy it to the connected device."*
>
> *"Create a new C++ application project using the installed SDK."*
>
> *"Import the GStreamer sample project from my SDK."*
>
> *"Register my existing project at \~/myapp with the installed SDK."*

> [card] [C++ Code Search](#c-code-search-rag)
>
> *"Find C++ sample code for initializing a camera pipeline."*
>
> *"Show me examples of camera sensor initialization from the SDK samples."*
>
> *"How do I set up a GStreamer source element? Find relevant samples."*

> [card] [Documentation](#documentation-search)
>
> *"How do I flash a device using the extension?"*
>
> *"What are the prerequisites for building a project?"*
>
> *"Explain how SSH key-based authentication is set up for a device."*

> [card] [AI Models](#ai-models)
>
> *"Download the Llama 3 model for my RB3 Gen 2 device."*
>
> *"Profile the ResNet50 model on my connected device and show me the results."*
>
> *"Compile the Whisper model for my device and push it once done."*
>
> *"What AI models are available for the RB5 platform?"*

> [card] [SDKs & OS Images](#sdks-and-os-images)
>
> *"What SDKs are available for my devkit?"*
>
> *"Download the latest OS image and SDK for my RB3 Gen 2 in one go."*
>
> *"Is my SDK download finished yet?"*

> [card] [Skills Hub](#skills-hub)
>
> *"What skills are available for my connected device?"*
>
> *"Install the best skill for working with GStreamer pipelines."*
>
> *"Show me the documentation for the camera-capture skill."*

> [card] [Overlay Configuration](#overlay-configuration)
>
> *"What overlay configurations are available on my device?"*
>
> *"Switch my device to the upstream overlay configuration."*

> [card] [Delta Flash](#delta-flash-incremental-deploy)
>
> *"Show me which files would change if I did an incremental deploy now."*
>
> *"Sync only the changed build artifacts to my device over SSH."*
>
> *"Do a dry run of the delta flash and then apply it if it looks right."*

***

### Device Management

These tools give AI assistants full visibility and control over connected Qualcomm hardware. Most other tool categories depend on an active device being selected first.

> [accordion] list\_devices — List all connected devices
>
> Returns every device currently detected by the extension, including serial
> numbers, connection type (SSH or EDL), and device state.

> [accordion] get\_active\_device — Get the selected device
>
> Returns details about the device currently set as active. Most other device
> and project tools operate on this device implicitly.

> [accordion] select\_device — Set the active device
>
> Sets a device as active by serial number. If the serial number is omitted or
> not found, an interactive VS Code picker is shown.

> [accordion] add\_device\_via\_ssh — Add a device over SSH
>
> Adds a new device to the known list using an SSH connection. Can generate a
> new SSH key pair; prompts securely for the device password when needed.

> [accordion] validate\_ssh\_connection — Check SSH reachability
>
> Verifies that the SSH connection to a device is reachable without performing
> any other operations.

> [accordion] update\_ssh\_config — Edit SSH configuration
>
> Updates the SSH connection parameters (host, port, user, key) for the active
> device.

> [accordion] list\_wifi\_networks / connect\_wifi / disconnect\_wifi / get\_wifi\_status — Wi-Fi management
>
> Scans, connects, disconnects, and inspects Wi-Fi on connected devices over
> SSH.

> [accordion] reboot\_device — Reboot a device
>
> Reboots the active device normally or into EDL (Emergency Download) mode.
> EDL reboots require explicit destructive confirmation.

> [accordion] flash\_device\_software — Flash an OS image (Linux-only)
>
> Flashes an OS image onto the active device. Streams real-time stage
> notifications: EDL wait, flashing progress, and CamX overlay pending status.

> [accordion] apply\_camx\_overlay — Apply CamX overlay (Linux-only)
>
> Applies the CamX proprietary overlay after `flash_device_software` signals
> it is awaiting the overlay step. Requires Qualcomm Linux Proprietary v2.0+.

> [accordion] list\_devkit\_catalog — Browse supported devkits
>
> Returns all supported Qualcomm devkits with IDs, chipsets, supported OS
> versions, and host architectures.

> [accordion] get\_pre\_flash\_data — Pre-flash instructions
>
> Retrieves the pre-flash instructions URL for a given devkit and target OS
> combination.

> [accordion] get\_auto\_activate\_status / update\_auto\_activate — Auto-activation control
>
> Reads and sets whether the extension automatically selects a device when one
> is connected.

> [accordion] reset\_device\_selection — Clear the active device
>
> Clears the active device selection and resets the auto-activation state.

> [accordion] remove\_device — Remove a known device
>
> Removes a device from the known device list. Requires destructive
> confirmation.

> [accordion] provision\_ufs / program\_cdt — Low-level provisioning (Linux-only)
>
> **provision\_ufs**: Provisions Universal Flash Storage on the device in EDL
> mode. **program\_cdt**: Programs the Configuration Data Table. Both require
> EDL mode and destructive confirmation.

***

### Projects

Build, deploy, and run applications directly from an AI conversation. The full workflow — from project creation to on-device execution — is available as tool calls.

> [accordion] build\_project — Build (Linux-only)
>
> Runs the `qvsce-build` task. Supports a headless mode that accepts
> `projectPath` and `sdkId` directly — no open VS Code window required.

> [accordion] deploy\_project — Build and deploy (Linux-only)
>
> Builds the project and deploys the binary to the active device via
> SSH. Supports headless mode.

> [accordion] run\_project — Run on-device (Linux-only)
>
> Executes the deployed application on the active device. Supports headless
> mode.

> [accordion] import\_sample\_project — Import from SDK (Linux-only)
>
> Imports a sample project from an installed SDK into the VS Code workspace,
> including source and build dependencies.

> [accordion] create\_project — New project from template (Linux-only)
>
> Creates a new C/C++ Application or GStreamer Plugin Application from an SDK
> template, then configures `.vscode` with `tasks.json`, `launch.json`, and
> `c_cpp_properties.json`.

> [accordion] register\_project — Register an existing project
>
> Associates an existing local project with an installed SDK and writes the
> `.vscode` configuration files needed to build and debug it.

> [accordion] update\_preferred\_sdk — Switch the active SDK
>
> Sets or refreshes the SDK used by the workspace for compilation and
> deployment.

***

### C++ Code Search (RAG)

The extension maintains a local TF-IDF index of C++ code samples that AI assistants query automatically when writing or explaining code.

> [accordion] search\_cpp\_samples — Search for C++ code
>
> Searches the index using TF-IDF similarity against a natural-language or
> code query. Returns up to 10 function-level code chunks with relevance
> scores.

> [accordion] build\_cpp\_rag\_index — Build the search index
>
> Builds or rebuilds the C++ code search index from the directory set in
> `qvsce.cppRag.samplesDir`. A built-in default index is always present; user
> samples are merged on top. Returns the total chunk and vocabulary counts on
> completion.

> [accordion] get\_cpp\_rag\_stats — Index statistics
>
> Returns the total chunk count, vocabulary size, default chunk count, and
> user chunk count for the current index.

> [!NOTE]
>
> The index is built and stored entirely on your local machine — no source code
> or index data is transmitted externally. If multiple developers on a team need
> the same coding conventions reflected in their agent's suggestions, each
> developer must configure `qvsce.cppRag.samplesDir` and run
> `build_cpp_rag_index` individually. Teams can share the same source directory
> (for example, via a shared repository) to achieve consistent results, but the
> index build must be performed locally by each developer.

***

### Documentation Search

The assistant calls these tools automatically — on tool failures, before complex operations, and whenever you ask about IDE features.

> [accordion] search\_documentation — Search IDE docs
>
> Searches the Qualcomm DragonWing IDE documentation via the external docs
> server. Called automatically on tool failures, before complex multi-step
> operations, and whenever you ask about IDE features.

> [accordion] get\_documentation — Retrieve a specific page
>
> Fetches a documentation page or resource by URI from the DragonWing docs
> server.

***

### AI Models

Full integration with Qualcomm AI Hub for cloud profiling and compilation, plus local on-device benchmarking.

> [accordion] list\_ai\_model\_catalog — Browse the AI Hub catalog
>
> Lists AI models available for download, with optional filters for device
> name and model type (LLM, vision, audio, etc.).

> [accordion] download\_ai\_model — Download a model
>
> Downloads an AI model from the catalog. Fetches available precision/runtime
> variants first; an interactive picker appears if a variant is not specified.

> [accordion] get\_ai\_model\_status — Model status overview
>
> Returns the current download, compile, and flash status of all models in the
> local cache.

> [accordion] push\_ai\_model\_to\_device — Flash a model to hardware
>
> Pushes a downloaded model to the active device as a background operation.

> [accordion] profile\_ai\_model — Cloud profiling via AI Hub
>
> Submits a profiling job to Qualcomm AI Hub for a specified device. Returns a
> job reference for polling.

> [accordion] local\_profile\_ai\_model — On-device profiling
>
> Runs a profiling job directly on the active device — measures real inference
> time and memory usage on hardware.

> [accordion] get\_ai\_model\_profile\_result / get\_ai\_model\_local\_profile\_result — Retrieve results
>
> Polls profiling job results keyed by model UID — cloud or on-device.

> [accordion] compile\_ai\_model — Compile and customize a model
>
> Compiles a model via AI Hub for a specific device target, with optional
> precision and optimization settings.

> [accordion] get\_ai\_model\_benchmark — Benchmarking data
>
> Returns inference time, memory usage, layer counts, and tool versions for a
> model. Filterable by device, precision, and runtime.

> [accordion] import\_ai\_model — Import a local model file
>
> Imports a model file already on disk into the extension's model cache.

> [accordion] remove\_ai\_model — Remove a model
>
> Removes a previously downloaded model from the local cache. Requires
> destructive confirmation.

> [accordion] list\_ai\_model\_device\_types — Supported device types
>
> Returns all device types that support AI Hub profiling and compilation jobs.

***

### SDKs and OS Images

Typically used once during environment setup — browse catalogs, download, and install the SDK and OS image for your devkit.

> [accordion] download\_devkit\_software — Bundle download
>
> Downloads both an OS image and an SDK in a single tool call. Either
> component can be omitted.

> [accordion] list\_sdk\_catalog — Browse available SDKs
>
> Lists SDKs for a given devkit and OS version: names, IDs, host
> architectures, versions, and download URLs.

> [accordion] download\_sdk / get\_sdk\_status / install\_sdk — SDK lifecycle
>
> **download\_sdk**: Starts a background SDK download. **get\_sdk\_status**:
> Polls progress and install path. **install\_sdk**: Installs a downloaded SDK
> (Linux-only).

> [accordion] list\_os\_image\_catalog — Browse OS images
>
> Lists OS images for a devkit: names, IDs, versions, and download metadata.

> [accordion] download\_os\_image / get\_os\_image\_status — OS image download
>
> **download\_os\_image**: Starts a background OS image download.
> **get\_os\_image\_status**: Polls progress and flags CamX overlay requirements.

***

### Skills Hub

Browse and install agent skills that extend what AI assistants can do with the extension.

> [accordion] search\_skills — Find relevant skills
>
> Searches skills by name, description, and tags using weighted matching.
> Automatically filters by the connected device's environment.

> [accordion] download\_skill — Install a skill
>
> Installs a skill into the selected agent's user skills directory (Claude,
> Cursor, or Codex).

> [accordion] get\_skill\_documentation — Read a skill's SKILL.md
>
> Returns the full documentation for a single catalog skill, defaulting to the
> version best matched for the connected environment.

> [accordion] list\_skills — Full catalog
>
> Lists every skill from all configured repositories, including all published
> versions and per-version compatibility metadata.

> [accordion] delete\_skill — Uninstall a skill
>
> Removes a previously installed skill from the agent's skills directory.

> [accordion] refresh\_skills\_catalog — Force catalog refresh
>
> Bypasses the catalog TTL and re-fetches all configured skill repositories
> immediately.

***

### Overlay Configuration

Advanced configuration for Qualcomm Linux devices — inspect and switch device overlays over SSH.

> [accordion] get\_overlay\_configuration — Read overlay options
>
> Returns available overlay categories, the currently active overlay, and
> upstream/downstream options for the active device. Qualcomm Linux only.

> [accordion] apply\_overlay\_configuration — Apply overlay selections
>
> Applies a chosen set of overlay selections to the active device via SSH.
> Qualcomm Linux only.

***

### Delta Flash (Incremental Deploy)

Incrementally sync build artifacts to a device over SSH — faster than a full reflash when only a few files changed.

> [accordion] preview\_delta\_flash — Dry-run rsync
>
> Runs `rsync --dry-run` to report exactly which files would be transferred.
> No changes are made to the device. Linux and macOS only.

> [accordion] apply\_delta\_flash — Apply incremental deploy
>
> Executes the real rsync over SSH. Requires destructive confirmation.
> Optionally runs a post-command (e.g., `depmod`) and reboots the device.
> Linux and macOS only.

***

## Troubleshooting

> [accordion] MCP Server shows 'Stopped' in the sidebar
>
> 1. Open VS Code Settings and confirm `qvsce.mcpServer.enabled` is checked.
> 2. Check that nothing else is using port 3100 (`lsof -i :3100` on Linux/macOS). If there is a conflict, change `qvsce.mcpServer.port` to a free port.
> 3. Reload the VS Code window (`Ctrl+Shift+P` → **Developer: Reload Window**).

> [accordion] AI assistant cannot find any tools
>
> Confirm the MCP client is configured to connect to `http://127.0.0.1:<port>/mcp` (not `/sse` or the root path). The server uses the **Streamable HTTP** transport, not the legacy SSE-only transport.

> [accordion] search\_cpp\_samples returns no results
>
> The `qvsce.cppRag.samplesDir` setting may be empty or pointing to a directory with no `.cpp`/`.h` files. Set it to a valid path and call `build_cpp_rag_index` to populate the index.

> [accordion] Linux-only tools fail on WSL
>
> Make sure you have opened VS Code connected to WSL (the status bar should show **WSL: Ubuntu** or similar) rather than using a Windows-native VS Code window with the extension installed on the Windows side.
