The current source branch requires exactly MTPLX 2.10.2 on Apple Silicon macOS. The installer and runtime reject other MTPLX versions. There are no supported prebuilt Switchboard binaries.

Use GitHub's private vulnerability reporting for this repository. Do not attach prompts, model outputs, configuration files, logs, usernames, home-directory paths, API credentials, or model files to a public issue.

Include the smallest reproduction that establishes the problem, the affected source revision, macOS version, and MTPLX version. Redact local paths and session identifiers.

Switchboard launches a separately installed MTPLX executable and exposes a bounded, unauthenticated HTTP/1.1 gateway on each configured loopback endpoint. The gateway and private upstream daemons bind only to 127.0.0.1. Other local processes can send requests and, when load-on-request is enabled, cause a configured model to load. Treat the local account, selected MTPLX executable, model repositories, and model files as trusted inputs.

Gateway routing is fixed per endpoint, with bounded headers, request buffering, connections and admission queues. It never accepts a caller-selected upstream host or automatically replays a generation request. Public MTPLX mutation endpoints are blocked. Private upstream ports are unsupported client interfaces: activity checks reduce risk, but cannot atomically drain callers that bypass the gateway. Normal unload/reload/quit waits for accepted gateway work and fresh owned-daemon idle evidence; force-quit and outside process manipulation remain exceptions.

Model downloads are explicit and are passed as literal process arguments. Switchboard checks the selected repository identity, canonical managed-cache path, and MTPLX runtime contract before registration. It does not independently verify model-file signatures or hashes.