YawLabs/tailscale-mcp v0.23.1 on GitHub

The oam floor moves from 0.18.0 to 0.19.1, so the tailscale-mcp launcher no longer runs the server on oam 0.18.x or 0.19.0. 0.19.1 is verified on the installed binary (oam 0.19.1) through the launcher with OAM_BIN pointing at it and TAILSCALE_MCP_RUNTIME=oam: a full MCP handshake listing all 98 tools and all 4 resources, tailscale_tool_groups answering, a call with invalid arguments returning the same error, and the whole session byte-identical to the same session on Node, with no non-JSON line on stdout. The same session answers from oam run bin/tailscale-mcp.mjs (in-process), from oam run src/index.ts (no build step), and with TAILSCALE_MCP_SANDBOX=1. An oam below 0.19.1 is handled as any below-floor oam: Node under auto, an error under TAILSCALE_MCP_RUNTIME=oam. Run oam self-update.
TAILSCALE_READONLY, TAILSCALE_LOCAL_CLI, TAILSCALE_REQUIRE_APPROVAL, TAILSCALE_DEBUG and TAILSCALE_MCP_SANDBOX accept 1 or true in any case, with surrounding whitespace ignored. cmd.exe's set TAILSCALE_READONLY=1 && ... stores "1 ", which the old exact compare read as off -- the security opt-in failed open without a word. TAILSCALE_READONLY and TAILSCALE_MCP_SANDBOX now print one stderr warning when set to any other non-empty value (yes, on), since the guard is then off. TAILSCALE_MCP_RUNTIME is trimmed too, so "node " selects Node instead of being warned about and treated as auto.

The local-CLI tools pass windowsHide: true when they run the tailscale binary, as the launcher already did for its own spawns, so a server under a GUI client on Windows no longer flashes a console window per call.
scripts/check-oam-floor.mjs runs from a junctioned or symlinked checkout. It compared process.argv[1] with its own module URL, which Node realpaths and argv[1] is not, so through a link it never ran its checks and exited 0 in silence. Both sides are now realpath'd. Release tooling only.