Quick Start Guide
UnfoldAI is the safety and recovery layer for AI coding agents. It gives you a chat-first Command Center, local Activity receipts, Safety cards, deterministic validation, checkpoints, MCP-lite tools, and verification/recovery prompts.
Step 1: Open Command Center
Click the UnfoldAI icon in the VS Code Activity Bar.
When v5 Safe Agent Mode is enabled, Command Center is the primary route. It shows:
- local/privacy status
- latest workspace/session status
- top active risk, if any
- suggested local actions
- the existing chat composer
Step 2: Run Your AI Agent
Use Cursor, Claude Code, Codex, GitHub Copilot, Cline, Windsurf, or a terminal agent as usual.
UnfoldAI watches local workspace evidence and creates receipts when there is enough activity to summarize.
Step 3: Validate The Latest Session
In Command Center, click Validate latest session.
Validation is deterministic and local. It checks available evidence such as verification status, diagnostics summaries, dependency/config changes, protected-path signals, and capture quality.
Validation does not prove the code is correct. It tells you what the session evidence can and cannot confirm.
Step 4: Review Safety
Open Safety when UnfoldAI finds high-confidence risk cards.
Safety cards can point to:
- protected or sensitive areas touched
- dependency changes without successful verification
- verification gaps
- new diagnostics near changed files
- agent-loop patterns
- recurring local risk
From a card you can review evidence, validate, mark safe, dismiss, or generate a verification/recovery prompt.
Step 5: Use Prompts Carefully
Verification prompts ask an agent to verify before you trust the session.
Recovery prompts ask an agent to plan first, make the smallest safe change, run verification afterward, and ask before editing protected paths.
Prompts are copy-only. UnfoldAI does not auto-send them, auto-run commands, auto-apply fixes, or restore checkpoints.
Step 6: Configure MCP-lite If Needed
Open Command Center's MCP details and copy the manual MCP configuration for your agent.
MCP-lite tools are metadata-only by default. Agents can request checkpoints, list checkpoints, get safe diff summaries, inspect risks, and request verification or recovery prompts. Restore requests never restore directly.