A local-first control layer for tool-using AI agents. Wardrail is being built to enforce permissions before execution—and leave clear evidence of what happened.
For AI infrastructure engineers and security teams.
Our engineering commitments
Fail closed on uncertain authorization
Local-first, no required cloud inference
Evidence before security claims
The control layer
Conversational traces tell you what an agent requested. A protected execution boundary helps you control what reaches the tool.
Illustrative policy / repository agent
workspace.readALLOW
protected.pushAPPROVAL
production.deleteDENY
No decision → no protected execution.
Register an agent, define its permitted resources, and wrap supported tools. Deterministic policy checks are designed to authorize each protected action before execution—not ask a model whether it feels safe.
Identity → policy → protected execution
Follow tool attempts, policy decisions, execution results, and failures in one correlated timeline. Hash-linked events support integrity verification against trusted checkpoints.
Observed evidence stays separate from inference
Run versioned attack and normal-behavior fixtures locally. Compare against a preserved baseline, with security failures, false denials, and infrastructure errors reported separately.
Reproducible scenarios. Explicit failure categories.
No universal protection claims. No automatic rollback promises. Wardrail’s scope is the execution boundary it can actually observe and enforce.
Our approachDeveloper first
The planned developer offering includes free local SDK and evaluation tooling. Paid team capabilities are under validation; availability and terms will be confirmed before release.
For individuals getting started
$0/month
Build with boundaries
Help shape Wardrail’s first release for real developer-agent workflows. Join early access to follow development.