ninja-p2pOpen a local conversation with an OpenAI agent to discuss a task, speak through your microphone, or show a selected camera or screen view. Keep task decisions and worker progress in the mailbox task page.
Set OPENAI_API_KEY in your local terminal, then run:
ninja-p2p-callTo include a task you can access:
ninja-p2p-call --config /private/mailbox.json --task TASK_IDFrom a checkout, build the CLI and use node runtime/call-server.mjs with the
same options. Open the private loopback URL printed by the command on that
computer. The key stays in the local server process. Keep the terminal running
while using the page; after a reload, reopen the original URL in a new tab.
Start a conversation#
- Select Start voice session. You can type immediately and hear the reply.
- Choose Share microphone to speak. Select Stop microphone to stop capture.
- Choose Share camera or Share screen to send selected frames. The browser lets you choose the screen or window. Use Stop visual sharing when finished, or select one image file instead.
- Select Interrupt to stop the current response, or End session to close the connection and stop microphone and visual capture.
Microphone and visual sharing start off. The page shows which inputs are being sent to OpenAI. Camera and screen sharing send one frame every three seconds; they are selected snapshots, not a continuous video track to the model. Image uploads accept PNG, JPEG, or WebP files under 5 MB and resize them before sharing. Audio replies are AI generated. Use the audio player's controls if your browser blocks playback.
Bring in a task#
When launched with a task ID, the page loads a preview using that mailbox's credential. Select Include this task when I send a message to share its current goal, state, approval, and result with OpenAI. Refresh or preview it before sending. Unchecking the option stops adding task context to subsequent messages; it does not erase content already shared in that model session.
The conversation discusses your task. It does not update tasks, approve requests, send mailbox messages, or control a browser. Use the mailbox page for task decisions. For an agent that operates a shared remote browser, use Ninja MCP's browser companion.
Continue without WebRTC#
Choose Use HTTPS text instead of voice, or type after ending the voice session. This uses an ordinary HTTPS request and works when a WebRTC connection cannot be established. Each text request contains your current message and any explicitly selected task context; earlier voice transcripts and images are not automatically copied into it. Include the context you need in the message.
The voice connection uses WebRTC between your browser and OpenAI. It is separate from a VDO.Ninja peer room, so other room members do not automatically join or hear it. Hosted mailbox tasks remain available independently of the call.
Models and limits#
The defaults are gpt-realtime-2.1 for voice and vision and gpt-4.1-mini for
HTTPS text. Set NINJA_CALL_MODEL or NINJA_TEXT_MODEL in the server process to
choose a compatible model available to your OpenAI project.
Each call ends after five minutes. A server process permits one active call, up to six call attempts, and 30 HTTPS text requests. Restart it to begin another bounded session. Model usage is billed to your OpenAI account. The application does not record locally; OpenAI processes the inputs and generated responses.
The local server requires its private access URL and checks request origin. Keep it bound to loopback. For a hosted deployment, add user authentication, account-specific usage limits, and HTTPS before exposing session creation.
The adapter follows OpenAI's Realtime WebRTC interface and Responses API.