Prepare an operations report
Work from approved inputs and prepare a draft for review.
Boundary: saving to an agreed destination requires the matching permission.
AI agent control plane Local-first alpha
Run AI agents within clear boundaries.
Gevurah is a control plane for teams building AI workflows. Each AgentRun brings the request, its permissions and its execution record together, with human approval when an action requires it.
Mock executor by default. Public product API and signup are disabled.
AgentRunIllustrative example
- 01
Prepare an operations report
Approved sample files only
- 02
Draft prepared for review
Waiting for approval
Save this report to the specified destination
- 04
Decision and outcome recorded
A flow illustration, not a live product session.
Mission Replay
Follow a request to its approval boundary.
Prepare a report from approved sample files, review the draft, then wait for permission to save that exact report. This is a local illustration. No tools run and no real action is approved.
Walkthrough ready. No action is being executed.
Request
Prepare a report from approved sample files.
Scope
Define the inputs and the permitted destination.
Prepare
Produce a draft for review.
Approval
Wait for permission to save this exact report.
Action
Continue only within the approved scope.
Record
Keep the decision and the recorded outcome.
An execution record keeps the decision with its outcome. A recorded outcome can be succeeded, failed or cancelled.
Open the full walkthrough ↗Example workflows
Put a boundary around the work that matters.
Start with a concrete task. Define the inputs, the output and the action that needs a person’s decision.
Review a proposed code change
Produce findings and a proposed change for a reviewer to inspect.
Boundary: changing production needs separate authorization.
Draft a customer response
Prepare a response using the context the agent is allowed to read.
Boundary: a draft is reviewed before any authorized sending step.
These are example workflows. Execution depends on the configured adapter and permissions.
The product
One run, a shared execution record.
AgentRun is the shared contract behind the product’s API, MCP, Desktop and Mobile design. The surfaces connect to one service, so permissions and state stay with the run.
The packaged product is a local-first alpha. Surface availability depends on the alpha profile and configured adapters. Public access is disabled.
APIMCPDesktopMobile
Explore AgentRun Scope & permissions
Define the resources and actions available to a run.
Review & approval
Wait when review or an exact action approval is required.
Execution record
Preserve the recorded outcome, including failure or cancellation.
Built from existing owner operations
The founder’s operating environment is the source of the evidence below. OpticoAI is built on Gevurah’s infrastructure and uses its technology. That operating experience guides the product being packaged today.
Owner operations and the packaged alpha are distinct. Their records do not establish customer-product reliability.
Operational evidence
A dated record. A defined method.
A fresh snapshot of the founder’s operating environment, measured on 26 September 2026. These counts describe recorded work and permission decisions.
5,518
Work-order files
Counted across the canonical queue and archives. Includes audits and reviews.
1,519
Result files
Markdown result files on disk. Includes unlinked records; this is not a success count.
893
Recorded policy decisions
Gate audit records dated 1 Aug to 25 Sep 2026. A permission decision is not proof of execution.
56BLOCK · permission denied
837ALLOW · permission granted
All 56 BLOCK records include a reason in this snapshot. ALLOW records permission, not successful execution or delivery.
How these figures were counted
Work-order files use the existing collector across the live queue and its two canonical archives. Results are Markdown files in the result directory. Policy figures come from parsing each gate audit row: 893 valid rows, no invalid rows, and a reason present in each of the 56 BLOCK records. The two file counts are not matched pairs and must not be divided into a completion rate.
No current completion percentage is published. Outcome parsing and linkage have known limitations, and recorded outcomes are self-reported. This snapshot is not an independent reliability benchmark.
Current boundaries
Product status and limits.
Local alpha
The product is being packaged for local use. The mock executor is the default; public signup and the product API are disabled.
Evidence scope
The figures describe the founder’s environment. They do not establish reliability for an external customer or production tenant isolation.
Human approval
The walkthrough is an illustration. It runs no tools, approves no real action and never sends a message.
Seed research
Gevurah Seed is a separate model research program. No Seed model serves production.
Talk to the founder
Which workflow needs clearer controls?
Describe the workflow, the tools involved and where approval is needed. Oren Oved builds Gevurah from hands-on operating experience.