Build a Secure Local AI Automation Agent That Plans, Approves, and Executes Tasks

Build a fully functional micro web app called Local Automation Agent Hub that simulates an OpenClaw-style automation agent.

The user describes an automation in natural language, and the agent must:
1. Convert it into a structured workflow
2. Propose concrete actions as a queued plan
3. Require explicit user approval before any “execution”
4. Execute actions via mock connectors (no real accounts required)
5. Produce a complete audit log and a replayable run history

This app must feel like an automation control center, not a chatbot.

⸻

Core Experience

User types something like:
“When an email contains the word invoice, extract the amount, save it to a local ledger, and remind me in 3 days.”

The agent outputs a workflow with:
• Trigger
• Filters
• Steps
• Outputs
• Risks
• Permissions needed

Then the user can “Run Simulation” to see the actions execute using mock data.

⸻

Required Features

1) Automation Builder (Natural Language → Workflow JSON)

Convert user text into a workflow format like:
• Trigger: Email Received (mock)
• Conditions: subject/body contains keywords
• Steps: Extract fields → Save record → Create reminder
• Outputs: ledger entry + reminder card

Render the workflow visually as a step-by-step pipeline.

⸻

2) Connector System (Mocked but Realistic)

Implement mock connectors (in-app only):
• Inbox Connector: generates sample emails
• Calendar/Reminders Connector: creates local reminders
• Ledger Connector: writes entries to a local table
• Webhook Connector: simulates POST requests and responses
• Notes Connector: saves notes/snippets locally

All connector actions must be logged.

⸻

3) Safety + Approval Gate (Non-negotiable)

Before execution, show an approval modal listing:
• Actions to run
• Data involved
• Risk level per action (Low/Medium/High)
• Required permissions

User must approve:
• Approve all
• Approve step-by-step
• Deny (and edit workflow)

No silent execution.

⸻

4) Agent Loop UI

The agent must visibly follow:

Interpret → Plan → Ask Approval → Execute → Verify → Summarize → Suggest Improvements

Show this as a timeline with statuses.

⸻

5) Audit Log + Run History

Store each run locally with:
• timestamp
• workflow JSON version
• inputs used
• steps executed
• outputs produced
• errors and retries

Allow “Replay Run” with same inputs.

⸻

Unique Twist

Add “Policy Rules” the user can set, such as:
• Never delete anything
• Never send outbound messages
• Never run high-risk actions without step approval
• Only allow saving data locally

The agent must enforce these rules when generating plans.

⸻

UI Requirements

Dark mode only, clean operator dashboard.

Layout:

Left: Automation prompt + policy rules
Center: Workflow pipeline (editable steps)
Right: Execution queue + audit log

Include:
• Step cards with icons
• Risk badges
• Run button with safe states
• Export workflow JSON button

⸻

Technical Requirements
• Single-page app
• Pure HTML/CSS/JS
• No external libraries
• LocalStorage persistence for workflows + run history
• Modular architecture:
• workflow parser
• policy engine
• connector runtime
• logger

Include a mock AI layer (a function) that can be swapped for real API later.

⸻

Internal AI Prompt Template

You are an automation agent.

Convert the user request into a safe workflow with:
• Trigger
• Conditions
• Step-by-step actions
• Outputs
• Risk level per step
• Required permissions
• Notes on safety

Follow policy rules strictly. Do not include actions that violate rules.

User request:
{{USER_TEXT}}

Policy rules:
{{POLICY_RULES}}

Return structured JSON only.

⸻

Output Requirements

Return:
1. Complete runnable app code
2. Explanation of workflow JSON format
3. Explanation of policy enforcement
4. Notes on swapping mock AI for a real API safely

No placeholders. No pseudo-code.