Nvidia Broadcast

NVIDIA Broadcast

Hard rules

1. Launching: start Broadcast only by running powershell.exe -NoProfile -ExecutionPolicy Bypass -File "<skill_dir>\scripts\launch_broadcast.ps1" with the terminal tool. When running inside WSL, follow Launch and version-check through interop in references/connection.md instead. Except for the security-policy fallback below and that WSL interop procedure, never invoke NVIDIA Broadcast.exe directly — not via terminal, execute_code, subprocess, or Start-Process.

   Security-policy fallback: use this only if security policy blocks the launcher script. Confirm that NVIDIA Broadcast.exe is not already running, then run: powershell.exe -NoProfile -Command "Start-Process -FilePath ([Environment]::ExpandEnvironmentVariables('%ProgramFiles%\NVIDIA Corporation\NVIDIA Broadcast\NVIDIA Broadcast.exe')) -ArgumentList '--launch-hidden' -WindowStyle Hidden" with the terminal tool.

2. Command tools: follow Which tool to run commands with below.

3. Never read effect state from disk — not AppSetting.json, not the registry, not logs, not the UI. Those are saved settings, not live state. Exception: gateway.json may be read and decoded only to obtain the gateway port for rule 4; it is connection configuration, not effect state.

4. Connecting to the gateway — exactly these steps, no other path:

   1. Before launching or registering the gateway, check the standard executable path and read its ProductVersion without launching it. If the executable is missing or its version is 2.2.x or older, stop the connection sequence and follow Installing or updating NVIDIA Broadcast. Do not return the restart-only response below.
   2. Check Broadcast is running: tasklist /FI "IMAGENAME eq NVIDIA Broadcast.exe" via terminal. If Broadcast is installed but not running, launch it immediately (rule 1); do not ask the user to start it manually. A gateway.json left from an earlier run proves nothing about whether the gateway is live.
   3. Whether already running or just launched, retry for up to 60 seconds: reread and base64-decode %APPDATA%\nvidia-broadcast\gateway.json → {"port": <number>}. Use the port from the latest successful read. Decode gateway.json within the command and output only the port. Never print its base64 contents because credential redaction may mask them.
   4. Register http://127.0.0.1:<port>/gateway with your MCP client as a Streamable HTTP server. You are the MCP client — register it yourself.
   5. Verify the server reports its name as nvidia-broadcast via the MCP client's server info. If it does not, treat NVIDIA Broadcast as unavailable, use the restart response below, and stop.
   6. Check whether the NVIDIA Broadcast MCP tools required for the user's request are available.

   If the required tools are available, the connection is ready; proceed with the requested operation.

   Stop if any of these occurs — report and go no further:

   • Registration fails, or
   • the server name is not nvidia-broadcast, or
   • after registering, any required NVIDIA Broadcast MCP tool is not in your tool list.

   Reply only: "NVIDIA Broadcast MCP tools are unavailable in this session. Please restart your MCP client and retry." Then stop. Do not add rationale, gateway URLs, port details, configuration instructions, or post-restart steps. Do not substitute another source or guess which effects are on.

5. Only after completing rule 4, if any required NVIDIA Broadcast MCP tool is still not in your tool list, that is the final answer. Stop. Its initial absence is not a reason to skip launch or registration. Do not try to make it work. Specifically, never:

   • Write, generate, or run any code that speaks MCP or HTTP to the gateway — urllib, http.client, socket, requests, httpx, curl, Invoke-WebRequest, or anything else. A correct, spec-compliant client is still a violation. The prohibition is on you speaking the protocol at all, not on which library you choose.
   • Search for, install, or run an MCP client binary — which mcp, mcp exec, npx @modelcontextprotocol/cli, pip install mcp[cli].
   • Scan ports, hunt the filesystem for an endpoint, or delegate to another agent or session. Use exactly the restart response from rule 4, then stop. A state read obtained by any other route is not a valid substitute, and a guessed effect list is simply wrong.

6. Recover from MCP connection failures. If any NVIDIA Broadcast tool call fails because the MCP connection was closed, refused, reset, disconnected, or otherwise became unavailable, repeat the connection sequence in rule 4:

   1. Check whether NVIDIA Broadcast.exe is running.
   2. If it is not running, launch it using rule 1. If it is already running, do not launch another instance.
   3. Reread gateway.json, obtain the current port, and register the gateway again.
   4. Verify the server reports its name as nvidia-broadcast. If it does not, use the short restart response from rule 4 and stop.
   5. Check whether the NVIDIA Broadcast MCP tools required for the user's request are available.

   If the tools are still unavailable, use the short restart response from rule 4 and stop. After reconnecting, retry a read-only or idempotent tool call at most once. Do not automatically repeat a non-idempotent request when the connection closed after submission because it may already have been accepted.

   Do not run this recovery sequence for normal gateway or tool errors such as invalid arguments, ambiguous_frame_rate, device_not_selected, file_not_found, rate limiting, or an effect operation failure.

Purpose

NVIDIA Broadcast applies AI effects to your camera, microphone, and speaker, and can apply the same effects to local media files. You drive it over a local MCP gateway that NVIDIA Broadcast runs on loopback. It also lets you check and change AI effects, select Studio Voice microphone profiles, manage camera, microphone, and speaker devices, and change camera resolution.

Prerequisites

• Windows host with NVIDIA Broadcast installed. 2.2.x and older have no MCP gateway and cannot be driven by this skill at all. When it is missing or too old, you may offer to install it from NVIDIA's update service — see Installing or updating NVIDIA Broadcast below. Never install without asking.
• Standard executable path: %ProgramFiles%\NVIDIA Corporation\NVIDIA Broadcast\NVIDIA Broadcast.exe.
• MCP gateway config readable at %APPDATA%\nvidia-broadcast\gateway.json.
• The current agent runtime acts as the MCP client and connects through Streamable HTTP over loopback.
• Running inside WSL: the gateway lives on the Windows side, so both the endpoint and gateway.json are reached differently — see If you are running inside WSL below.

Inputs

• set_effects: either effects[] with 1-32 unique effect IDs (each effectId + enabled, optional params), or action: "restore_previous_live".
• submit_file_processing: _version: 1, absolute inputPath, 1-32 unique effects[]; optional batchId, outputFolder, outputPath. There is no overwrite input — a fresh collision is rejected, while resume/retry replaces only that job's own partial output.
• list_devices: optional kind. set_active_device: kind + fresh deviceId, plus optional camera width, height, frameRate. set_camera_resolution: fresh camera deviceId, width, height, optional frameRate.
• Do not reuse device IDs from earlier sessions, and do not guess frame rates. Supply frameRate only when the user explicitly specified or selected it; the camera's current frame rate is not the user's choice for a new request.

Which tool to run commands with

Use terminal for shell commands and execute_code for Python. Never call browser_exec — its code parameter makes it look like execute_code, but it runs that code inside a browser session and opens Chrome. No task in this skill uses a browser.

Connecting

MCP client identity

In this skill, “MCP client” means the current agent runtime executing this skill—not a separate application, CLI, profile, agent, or session. This definition applies throughout the skill.

Do not conclude that no MCP client exists merely because Broadcast tools were not preloaded. Do not search client configuration files, invoke another client such as Claude Code, use a browser, browser-based MCP connection, browser automation, or browser-based remote debugging, delegate to another agent or session, or ask whether another MCP client is installed.

Streamable HTTP is the default and recommended transport. Register the gateway with your MCP client and let the client speak the protocol. Do not hand-roll HTTP requests — it is a normal Streamable HTTP MCP server, and your client already handles protocol version and transport. Add it to your MCP client configuration:

{
  "mcpServers": {
    "nvidia-broadcast": {
      "type": "http",
      "url": "http://127.0.0.1:18100/gateway"
    }
  }
}

Or, for clients that register servers from the command line:

<mcp-client> mcp add --transport http nvidia-broadcast http://127.0.0.1:<port>/gateway

Substitute the client's own command for <mcp-client> and the resolved gateway port for <port>; subcommand and flag names vary by client, but the three values being registered are the same ones in the entry above.

18100 is the default port and is correct on almost every install. No API key or header is required for loopback MCP clients. This URL and port are for the client's configuration, not for connecting directly.

If the default port does not work

The gateway takes the first free port in 18100-18109 and publishes the one it actually bound in %APPDATA%\nvidia-broadcast\gateway.json — base64, decoding to { "port": <number> }. Re-read that file on every retry rather than caching the port; a restart can move it, which is the one case where a registered URL goes stale.

1. Before launching or registering the gateway, read the executable's ProductVersion without launching it. If the executable is not there, NVIDIA Broadcast is not installed. If it is 2.2.x or older, that build has no MCP gateway and no amount of retrying will produce one. Either way, stop the connection sequence, tell the user what you found, and offer to install or update it — see Installing or updating NVIDIA Broadcast below. If they decline, or if installing is genuinely impossible from here — no Windows interop, no PowerShell — point them to https://www.nvidia.com/broadcast-app/ and stop. An earlier failed attempt is not one of those reasons, and neither is a download that had to be retried.

2. Check whether NVIDIA Broadcast.exe is running. If it is not, launch it by running the bundled launcher script — the primary launch command. Use the security-policy fallback in rule 1 under Hard rules only if policy blocks the script:

   powershell.exe -NoProfile -ExecutionPolicy Bypass -File "<skill_dir>\scripts\launch_broadcast.ps1"

   <skill_dir> is the absolute path in the skill_dir field of the skill_view result you already have. Run it with the terminal tool. Never with browser_exec — it takes a code parameter just like execute_code, but it runs that code inside a browser session and will open Chrome. Nothing in this skill ever needs a browser. Except for the security-policy fallback in rule 1 under Hard rules, do not write your own launch command or invoke NVIDIA Broadcast.exe directly with subprocess.run, check_output, Start-Process -Wait, or any command that waits for or captures output from the Broadcast process. Never launch a second copy when the process is already running.

3. Whether Broadcast was already running or was just launched, retry for up to 60 seconds: reread and decode gateway.json, then register http://127.0.0.1:<port>/gateway through the MCP client. If nothing answers, use the restart response in hard rule 4 and stop.

   Discovery is exactly these two places and nothing else: gateway.json and ports 18100-18109. Do not search the filesystem for gateway or config files, and do not read environment variables looking for an endpoint or credential — the loopback gateway publishes its port only in gateway.json and needs no key. When both places come up empty, stop endpoint discovery.

   Probe only through your MCP client. Register the candidate URL and let the client connect; a failed registration is the probe result. Never reach the gateway from a shell: no command-line web clients, no throwaway request scripts, and no scanning the port range yourself — see Safety and scope for the full list of prohibited alternatives. If your client cannot register MCP servers at runtime, follow If your client cannot register MCP servers below. Not being able to probe never licenses shell tooling as a substitute.

If your client cannot register MCP servers at runtime

Some clients only load MCP servers from a config file at startup. That is a client configuration problem, not a Broadcast problem, and it is never a reason to fall back to "just turn it on in the app yourself" without checking anything. Do the version check in step 1 first — it reads the filesystem and needs no MCP client — then report whichever of these you actually found:

• Broadcast is missing, or is 2.2.x or older. Offer to install or update it exactly as below. Installing needs no MCP client, and it is worth doing before any config change.
• A 2.3+ build is installed. Nothing is wrong with Broadcast. Tell the user to add the gateway to their MCP client's configuration and restart the client, quoting the entry from Connecting with the port you resolved. Say plainly that you cannot drive Broadcast until they do.

Name which of the two it is. "The gateway isn't available to this session" on its own leaves the user unable to tell whether the app is missing, too old, or merely unconfigured — and sends them off to click through the UI when one config line, or an install, would have fixed it for good.

If you are running inside WSL

The gateway runs in the Windows Broadcast process, so before registering anything establish that 127.0.0.1 in the distro really is Windows loopback, and read the port through the mounted Windows drive. Full procedure in references/connection.md; in short:

1. Detect WSL from /proc/sys/kernel/osrelease — it contains microsoft or WSL. Do not read environment variables to decide this.
2. Check wslinfo --networking-mode. On mirrored — or a WSL 1 distro — loopback is shared, so register http://127.0.0.1:<port>/gateway exactly as on Windows. On nat (the default) the gateway is unreachable from the distro: stop, tell the user, and offer mirrored networking, an MCP client running on Windows, or a tunnel terminating on Windows loopback. Never substitute another address for loopback — not the nameserver in /etc/resolv.conf, the default route from ip route, nor $(hostname).local; nothing is published off loopback and the gateway rejects every non-loopback Host by design.
3. Go through Windows interop for the rest: wslpath "$(cmd.exe /c 'echo %APPDATA%' | tr -d '\r')" to reach gateway.json, and powershell.exe -NoProfile -Command '...' to launch, version-check, and install. If the Windows drive is not mounted or powershell.exe is missing, say so and stop — do not search the Linux filesystem for gateway.json and do not look for a Linux-side substitute; ask the user to start Broadcast on Windows.

See references/connection.md for the version-check command, endpoint discovery, startup, access, WSL specifics, and rate limits.

On connect, the server's instructions already list the exact effectIds available right now, grouped by section, and the set_effects schema constrains effectId to that live set. So you can act immediately — you do not need a discovery call first.

Installing or updating NVIDIA Broadcast

Only when the version check above found no executable at the standard path, or a 2.2.x-or-older build. Never as a workaround for any other failure, and never as a side effect of an unrelated request. Full procedure, commands and failure modes in references/installation.md.

1. Ask what is available. GET https://ota.nvidia.com/release/available?product=rtxb&channel=OFFICIAL&version=$installed&cpuArchType=$arch, where $installed is the installed ProductVersion (0.0.0.0 when Broadcast is not installed) and $arch is aarch64 or x86_64, read from the registry as shown next — never chosen or defaulted. Substitute both values; never send the names themselves. Check both before sending — $installed a dotted version number, $arch exactly aarch64 or x86_64. The service never rejects a bad value: it answers [], which reads as "nothing newer". Read the architecture from the registry, not the process — $env:PROCESSOR_ARCHITECTURE says AMD64 inside x64 emulation on Windows on ARM, which fetches the wrong installer. Use $native = (Get-ItemProperty "HKLM:\SYSTEM\CurrentControlSet\Control\Session Manager\Environment" -ErrorAction Stop).PROCESSOR_ARCHITECTURE, map AMD64 or x86→x86_64 and ARM64 or ARM→aarch64 into $arch, and throw on anything else — never default to one. Those two values are the only ones the service accepts, and each serves a different installer. The reply is a list of applicable builds, newest first and already filtered to versions above the one you sent — take the first element. An empty list means there is nothing to install; say so and stop, it is not a failure to retry. Distinguish that from a call that never completed, which is a network error and is retryable.
2. Ask the user, and wait. Name the build and size, say it comes from ota-downloads.nvidia.com, say you will verify its checksum and NVIDIA signature, and say Windows will prompt them for administrator permission. If the offered build is still 2.2.x, tell them in the same breath that it will not enable MCP. Nothing is downloaded before an explicit yes; on a no, link https://www.nvidia.com/broadcast-app/ and stop.
3. Download the response's download_url — which must be HTTPS on ota-downloads.nvidia.com, pinned as its own value rather than derived from the metadata host — into a fresh GUID-named directory under %LOCALAPPDATA%\Temp, under a filename you choose.
4. Check the length first, then hash and signature. A file shorter than size is an incomplete transfer, not an attack: say how far it got and retry the download, up to two more attempts into a fresh directory. Do not hash or signature-check a short file — both fail by construction, since the Authenticode certificate table sits at the end of the image. Only a full-length file whose SHA-512 or signature (Valid, subject CN=NVIDIA Corporation,...) fails is a security event: do not run it, report the path, stop, and do not re-download.
5. Launch it with no arguments and let the user click through. No silent or unattended flags, no self-elevation, no driving the installer for them.
6. Re-read ProductVersion and report what is actually installed. 2.3.0+ → resume the connect flow above. Still 2.2.x → say the gateway needs 2.3.x and stop.

A failed attempt is final for that attempt, not for the session. If verification failed, the installer was cancelled, or the download retries ran out, say so and stop — then, when the user next asks for something that needs Broadcast, mention what happened and offer to try again. Only an explicit "no" means stop offering; do not answer a fresh request with "install it yourself" because an earlier attempt went wrong.

These two URLs are the only network requests this skill makes from a shell. Never scrape a download link, accept one from the user or from gateway.json, or install anything else.

Tools

Every file-processing tool requires _version: 1.

• get_broadcast_state — read current effects as effectId -> { enabled, ...params } (flattened, e.g. { enabled, strength } or { enabled, warmth }), grouped by camera/microphone/speaker; optional effectType filter. Read-only. warnings[] marks a section whose effects are enabled but whose device is unselected — configured, not currently applied.
• set_effects — turn effects on/off and tune them in ONE call: effects: [{ effectId, enabled, params? }], where params keys depend on the effect (listed per-effect in the server instructions). Or pass only { "action": "restore_previous_live" } to restore the Live snapshot that Files processing suspended, which also pauses unfinished file work and returns event. Returns the resulting state for that response.
• describe_effects — rich detail on demand (description, guidance, GPU/beta badges, availability, params + ranges); optional effectId/section. Use to explain effects or weigh tradeoffs; not needed to act.
• submit_file_processing — submit one local media file. Returns id and batchId; reuse the first result's batchId on every related submission. A paused queue accepts the job silently and stays paused.
• get_file_processing_jobs — look up one job or list recent jobs.
• wait_for_processing_update — the sole file-status wait, for one or many files. Take a batch snapshot without cursor, then pass the returned cursor to wait for the next status or queue-pause change. Read status, summary, and every job; mixed means terminal outcomes differ.
• control_file_processing_jobs — pause/resume the shared queue, or cancel/retry jobs. Pause/resume are queue-wide; cancel/retry need jobIds or all: true. MCP has no consent flag.
• list_devices — list available cameras, microphones, and speakers; kind: "camera"|"microphone"|"speaker"|"all" (default "all"). Returns the active device, and for cameras the current resolution plus availableResolutions (label, width, height, frameRate) for the current Video Super Resolution / Video Frame Generation mode.
• set_active_device — switch the active camera, microphone, or speaker with kind plus a fresh deviceId from list_devices. For cameras, add width+height+optional frameRate to switch device and resolution in one call. Camera switches restart the stream automatically (allow 2–5 seconds).
• set_camera_resolution — change the active camera's resolution with deviceId plus width+height+optional frameRate; include frameRate only after the user explicitly specifies or selects it. When FPS is unspecified, omit frameRate so the gateway can accept a unique mode or return the available choices for an ambiguous size.

Invalid requests — including unavailable or ambiguous resolutions — return a gateway tool error before anything changes. How to read the success-path fields (success, alreadyActive, alreadySet, frameRateMismatch, failed[], skipped[]) is under Instructions.

Effect ids and params mirror the app's UI (e.g. "Virtual key light" → virtual_key_light with warmth). The exact ids + each effect's params are in the server instructions. See references/effects.md.

For file processing, nest every effect's options under params, exactly as with set_effects. Frame Generation accepts multiplier 2|4 and model "performance"|"balanced"|"quality".

Studio Voice microphone profiles

Set the Studio Voice profile through params.micProfile in either set_effects or submit_file_processing. Read the capability-aware enum from the tool schema or server instructions; the current supported profiles are default, bright, full, and warm. Do not send flat, colored, or thin when they are absent from the advertised enum.

Examples

Complete tool calls — single and combined effects, per-effect params, background replace, frame generation, Studio Voice profiles, restore_previous_live, state reads, device switches and resolution changes — are in references/examples.md. Load it when you need the exact JSON shape.

Worked flows

Multi-step tasks have an expected shape. Follow it exactly.