Autonomy without an open bar
For autonomous validation without supervision, "can run any command" is a risk, not a feature. TecnicaiT is bounded by design across several layers.
Closed action vocabulary
The set of actions it can run is validated against a schema. Anything outside the vocabulary is rejected before it executes.
Protected paths in two layers
A hardcoded universal layer (engine, version control, secrets, dependencies) plus a per-client configurable layer. No finding can touch what is protected.
Hard budgets per procedure
Limits on actions, tokens and time per run. Exceeding them triggers a controlled abort, never an infinite loop or runaway cost.
Green tests as a mandatory gate
No applier change reaches a PR without passing the test suite. If tests stay red it reverts or retries with self-reflection — it never forces the commit.
Signed audit trail
Every run is cryptographically signed and tied to the versioned procedure that originated it. Reproducible traceability for certification evidence.
Your key, your model
BYOK by default: you bring your own AI provider key. Inference cost is yours and transparent, and the data flows through your account.