尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權,自行管理介紹,無需管理員核准。
登入並產生驗證檔案,在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案,即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入,將前往登入頁面。登入後請回到此處,再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify,即可驗證所有權並自行管理介紹。
一分鐘完成驗證Code-enforced capabilities for local AI agents — wallets, servers, payments inside walls the model can't override. scutbench grades which models can use them safely.
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Real-world skills. Hard limits.
scutl packages real-world capabilities for local AI agents — wallets,
servers, paid services, monitoring, backups — as code-enforced
recipes: the spending caps, allowlists, and teardown guarantees are
enforced by the installed CLI before any API call leaves the box,
not by prompt text the model can talk its way around.檢視完整 README ↗Real-world skills. Hard limits.
scutl packages real-world capabilities for local AI agents — wallets, servers, paid services, monitoring, backups — as code-enforced recipes: the spending caps, allowlists, and teardown guarantees are enforced by the installed CLI before any API call leaves the box, not by prompt text the model can talk its way around.
scutbench is the evidence layer: which local models can actually use those capabilities safely, graded on real tool-calling against mock providers that lie, time out, and try the tricks real counterparties try — before money, infrastructure, or identity is at risk.
ADAPT.md and it installs itselfgit clone https://github.com/murdarch/scutl
cd scutl
./tools/first-proof.sh http://localhost:8080 # your local OpenAI-compatible server
20 mocked scenarios, no model spend, no accounts — proves your harness + model can drive a recipe's tools before anything real is attached.
Each recipe is a molecule, and every part is in this repo:
| Part | What it is |
|---|---|
recipes/<id>/recipe.yaml | The manifest: decision tree, code-enforced walls, provider contracts, acceptance tests. The manifest is the recipe page on scutl.org. |
recipes/<id>/<cli>/ | The typed CLI component that enforces the walls (exit-code taxonomy is the protocol) |
recipes/<id>/ADAPT.md | Agent-facing installer, proven by fresh-agent acceptance runs (the pattern) |
scutbench/ | The mocked-twin benchmark deriving scenarios from each manifest's failure modes |
ladder/ | The live-rail run harness and receipts behind every Shipped badge |
site/ | The static site generator — pages render from manifests, so the site cannot drift from the recipes |
Recipes are grouped by exposure, and the intended path is to climb: disposable/testnet first, then read-only, then act-with-approval, then hard-capped unattended action, and only then real money and persistent infrastructure. Every claim on the site links its evidence: graded reports, on-chain transactions, signed checksums.
The site never claims more than the artifacts support: a shipped recipe without an installer says so, cells without a pinned environment record wear a visible partial-provenance marker, and the held-out benchmark rounds that keep the public numbers honest stay private.