scutl

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.

• Site / catalog: https://scutl.org — every recipe's manifest, rendered and served verbatim, checksummed and ssh-signed
• Benchmark grid: https://scutbench.scutl.org
• Agent front door: https://scutl.org/llms.txt — point your agent at a shipped recipe's ADAPT.md and it installs itself
• Two-minute proof (nothing at stake): https://scutl.org/start-here.html

Try it

git 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.

What's in a recipe

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

The trust ladder

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.

Status honesty rules

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.