Authorize exactly. Execute once. Verify the effect.
RuleOak v0.50.0 is an Apache-2.0, local-first Effect Boundary for supported AI-agent and MCP tool actions. It separates permission from outcome so a timeout or lost response does not automatically become a duplicate retry.
The problem
action dispatched
↓
external effect happens
↓
response is lost
↓
caller sees timeout
↓
blind retry can duplicate the effectOrdinary authorization answers whether an action may execute. RuleOak also models what happened after execution and whether the available evidence supports retry.
The chain
Start with the right page
Architecture
The authority/effect/evidence chain and integration boundaries.
Use cases
Filesystem, MCP, deployment, outbound actions and candidate-source review.
Concepts
Authority Envelope, ExecutionLease, EffectReceipt, Effect Profile and Flight Recorder.
Threat model
What RuleOak reduces, what it only partially addresses and what remains out of scope.
Examples
Supported integration patterns and the questions to answer before declaring a route protected.
Verify v0.50.0
Manifest, checksum, reproducibility, package and hosted attestation guidance.
Choose an operational path
Decision guide
Decide whether RuleOak fits the problem before integrating it.
Deployment patterns
SDK, MCP stdio, Streamable HTTP, REST and operator CLI boundaries.
Evidence model
Understand the difference between authorization, execution and effect evidence.
Safe retry
Handle timeout and response loss without blindly repeating the action.
MCP integration
Protect supported routes without confusing discovery with authority.
Node.js SDK
Guard application-owned functions and tool registries.
Project boundary
Secondary capability: AI-produced code review
The repository also contains an experimental Intent-Verified Development stack for reviewing external source candidates against declared intent and baseline-owned qualification. It is deliberately secondary to the runtime Effect Boundary and does not let a coding model certify itself.