vsync

The audit trail for AI coding agents.
Know what your AI decided, why it decided it, and whether it's safe to continue — without ceremony.

Quick Start · Zero-friction sessions · Tools · Commands · Trust & Security · Pricing · Docs

---

The problem

You're shipping code written by AI agents. But you can't answer basic questions:

• What did the AI decide and why?
• Did two agents just edit the same file with conflicting assumptions?
• Why did Cursor rebuild what Claude Code already decided?
• Is it safe to deploy what the AI wrote?

There's no git blame for AI decisions. Chat history disappears. Context doesn't follow you across tools.

The solution

vsync watches every AI coding session and produces a cryptographically signed trace: what was asked, what the AI decided, what changed, and why.

Open chat → vsync loads last session → You work → vsync captures → Idle → auto-save

One line: we make AI coding agents auditable — without ceremony.

Install the CLI from npm (@verifiedstate/sync). This repo is the public home for docs, examples, and templates.

---

Quick Start

npx @verifiedstate/sync init

One command (~30 seconds). It will:

1. Open your browser for one-click sign-in
2. Detect your AI tools (Claude Code, Cursor, Windsurf, Codex)
3. Write config files (CLAUDE.md, .cursor/rules/, AGENTS.md, MCP config)
4. Start a background daemon that captures every turn
5. Generate live hot-context for your repo

vsync start # start the daemon
vsync status # health + recent session activity
vsync stop # stop the daemon

---

Zero-friction sessions

You should never type "save session" or "load session." vsync handles the full lifecycle:

When	What happens	What you see
Open a new chat	vsync calls session_load, writes summary to .cursor/hot-context.md	macOS notification: ✦ VerifiedState · {project}
While you work	Every turn captured to your VerifiedState namespace	vsync status shows recent sessions
Go idle ~5 min	Auto-save: session handoff + audit summary	Dashboard session updated
Next chat	Auto-start again — context at the top of hot-context	No copy-paste, no MCP commands

Session summary lives at the top of .cursor/hot-context.md (and .cursor/vstate-session.md for full detail). Agents read it automatically.

Disable notifications: VSYNC_NO_NOTIFY=1 vsync start

---

Supported Tools

Tool	How it captures	Status
Cursor	Chat bubbles, agent transcripts, tool calls (edit_file_v2, bash, MCP, …)	Stable
Claude Code	Watches ~/.claude/projects/**/*.jsonl	Stable
Windsurf	Reads state.vscdb via sql.js	Stable
Codex CLI	Watches ~/.codex/sessions/**/*.jsonl	Stable
Git	Polls for new commits, captures metadata	Stable
Any MCP client	Via mcp.verifiedstate.ai	Stable

---

What It Captures

Data	Example
Session turns	User prompts + AI responses
Decisions	"Use batched updates to avoid long locks"
Tool actions	Read, Edit, Bash, MCP calls (truncated)
Code change hints	File paths, diff previews where available
Git metadata	Commit hash, message, branch, author
Agent identity	Tool, model, session ID

What it does NOT capture

Never collected
Full source code files (unless in AI chat)
Environment variables or .env files
API keys, secrets, tokens, credentials
Private keys (PEM/SSH patterns scrubbed)
Network traffic or browsing history

See SECURITY.md for the full list.

---

Commands

Command	Description
vsync init	Sign in, detect tools, write configs, start daemon
vsync start	Launch background daemon
vsync stop	Stop daemon (flushes session on exit)
vsync status	Daemon health + recent auto-started sessions
vsync logs	Tail recent daemon activity
vsync recap	Summarize recent memory activity
vsync agents	List agents that have written to memory
vsync why	Ask why a decision was made
vsync last	Show last N ingested items
vsync trail	Audit trail for an entity
vsync diff	What changed between two points in time
vsync handoff	Context packet for the next agent session
vsync audit	View local audit log (every API call logged)
vsync audit verify	Verify audit log hash chain integrity
vsync allow <path>	Add project to allow-list
vsync deny <path>	Remove project from allow-list
vsync start --dry-run	Print payload, skip network

Full CLI reference: verifiedstate.ai/docs

---

Trust & Security

vsync is designed so you don't have to trust us — you can verify.

Observe-only

vsync is a read-only observer. It cannot:

• Run shell commands (except git log for commit metadata)
• Write to your source files
• Modify git state (no commits, no pushes)
• Access files outside declared paths
• Send data anywhere except your configured endpoint

Verify for yourself

vsync start --dry-run --foreground # see exact payloads before send
vsync audit # every API call logged locally
vsync audit verify # hash chain integrity check
npm audit signatures # npm provenance (@verifiedstate/sync)

Secrets filtering

• .env files are hard-rejected — never read, period
• API keys (sk-*, ghp_*, AKIA*), bearer tokens, PEM headers are regex-filtered
• Redacted items logged locally: [redacted: pattern_name]

Full security details: SECURITY.md

---

What you see

The VerifiedState dashboard shows:

• Session handoff — summary, files changed, next steps, blockers
• AI Work Brief — what your agents did today, in plain English
• Trace view — why (prompt, decisions) + what changed (files, tools)
• Hot Files — files with the most AI edits, flagged for review
• Collision Detection — alerts when 2+ agents edit the same file
• Signed Receipts — cryptographic proof of what happened and when

---

Pricing

	Free	Pro	Team	Enterprise
Price	$0/mo	$19/mo (founder)	$59/seat/mo	Custom
For	Individual devs	Solo builders	Teams	Orgs
Session capture	✓	✓	✓	✓
Cross-tool memory	✓	✓	✓	✓
Signed receipts	✓	✓	✓	✓
Audit exports		✓	✓	✓
Shared team memory			✓	✓
SSO / compliance				✓

View full pricing →

Free is for individual developers and personal projects. Team or organizational use → Enterprise.

---

How it works

Your machine VerifiedState
+-----------------+ +------------------------+
| AI session files |--read-only--> | api.verifiedstate.ai |
| (Cursor, Claude | --scrub--> | /ingest |
| Code, Windsurf, | --send--> | | |
| Codex) | | signed receipt |
+-----------------+ +------------------------+
|
v
~/.vsync/audit.jsonl (local log of everything sent)

1. Daemon watches AI session files (read-only)
2. Payloads pass through secrets filter
3. Every API call logged to local audit file before sending
4. VerifiedState extracts decisions and creates signed receipts
5. Queryable via dashboard, CLI, or MCP tools

---

For AI agents

Add to your MCP config (Claude Code, Cursor, or any MCP client):

{
"mcpServers": {
"verifiedstate": {
"url": "https://mcp.verifiedstate.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_API_KEY" }
}
}
}

MCP tools include memory_query, session_load, session_save, working_state_get, and more. vsync auto-starts sessions — agents don't need to call session_load manually if the daemon is running.

Docs: verifiedstate.ai/docs · Cursor: verifiedstate.ai/docs/cursor

---

This repo

Path	Purpose
README.md	Product overview (you are here)
SECURITY.md	What leaves your machine
CHANGELOG.md	Public release notes
examples/	Before/after, generated config samples
templates/	What vsync init writes to your repo

CLI source ships via npm: @verifiedstate/sync

---

Contributing

Issues and feature requests: github.com/verifiedstate/vsync/issues

Security vulnerabilities: security@verifiedstate.ai

License

MIT — VerifiedState

---