Xnurta announced on November 19, 2026 that its MCP server now reads and writes: AI assistants like Claude and ChatGPT can bulk-adjust bids and budgets, change keywords and targeting, create campaigns, and archive them — from natural-language instructions. Counted from the launch release and Xnurta's MCP page: 3 modes (Read, Write, Optimize), 5 expanded capabilities, 3 Amazon formats, about 15 months of history. The only per-action guard named is a preview plus human confirmation for budget, bid and archive changes. Per-action decision scores named: 0 — and Xnurta's own FAQ notes archives can't be restored.
ScriptMasterLabs ran the shapes through its live decision harness on November 20, 2026 (local-heuristic-v1, calibrated=false; every value identical across two runs): the performance read auto-acted at 0.9545 (bounded scope) and 0.9054 (sentence form); every write — confirmed bulk adjust, and bulk adjust + archive without asking — held at 0.50 advisory; score forms escalated at 0.35. The harness's limits are stated plainly in the canonical: it held even the confirmed write, and its score form did not discriminate.
The graded layer is ScriptMasterLabs Verdict Protocol: $0.005 USDC per decision and $0.005 per seal on Base via x402 (verified live November 20, 2026 — /decide 402 quoting 5000 atomic USDC), 5 Jev jurors scoring a frozen capsule, the receipt sealed whether they approve or refuse.
Full measurement, comparison table, and the SELF-PROOF settlement record: https://scriptmasterlabs.com/xnurta-mcp-server
Catalog: https://scriptmasterlabs.com/sml-products
Top comments (0)