Turn a Local Agent Plan into a Safe Workstation Flow

Choose the goal before the tool

A useful local plan begins with the expected result, not a product name. A goal such as 'turn a customer list into shareable JSON' can be separated into sensitive-data review, masking, structural validation, and export. Local Agent matches goal signals to the catalog; it does not invent facts like a generative LLM.

Before accepting a plan, read why each step was selected and inspect alternatives for unsupported formats. Confidence is an explainable matching signal, not an accuracy guarantee.

• State the input type and final output.
• Separate consequential operations.
• Compare alternatives by rationale and data compatibility.

One-click handoff is not automatic execution

Sending a plan to Workstation writes a bounded document of tool IDs and edges to same-tab sessionStorage. Workstation builds nodes but never selects a file, enters a password, runs code, or starts a download.

The first node may receive the goal as input. Data reaches later nodes only after the user chooses a result handoff. This boundary prevents a weak plan from silently propagating through the chain.

• Read every node input before running it.
• Check output compatibility with the next tool.
• Undo or disconnect when a step fails.

Experiment safely with version history

Node, edge, and layout changes enter the undo history. Save the project locally before a major change, then work in small verifiable steps so failures remain easy to isolate.

Projects are AES-GCM encrypted in on-device IndexedDB, which does not make a compromised device or malicious extension safe. Keep sensitive input out of recipe URLs and compare P2P safety codes through a separate channel.