The problem hasn't gone away
Every AI coding session your team runs is a black box. Something happens, code changes, and there's no durable record of what the model was asked, what it decided, or what it touched. That gap is fine until it isn't — a compliance review, an incident post-mortem, or a subtle change that nobody remembers authorizing.
chron-mcp is an MCP server that creates a tamper-evident audit trail of every AI session, stored locally in a SQLite database on your own machine. No data leaves your environment.
Works today with Claude Code, Cursor, Windsurf, Codex, and any MCP-compatible client.
npx chron-mcp
What's shipped
Session verification — every session gets a cryptographic integrity check. You can prove a record hasn't been modified after the fact, which matters when presenting audit evidence to anyone else.
Export receipts with delivery states — when you export a session, chron now tracks the delivery: prepared, delivered, partially_delivered, failed, or unknown. The receipt stores a fingerprint, not the content. The UI says "collector accepted" deliberately — chron records the handoff, not what happens after.
Multi-client support — the same local database handles sessions from multiple AI tools side by side. Claude Code and Cursor in separate terminals, both landing in the same audit store, tagged by client.
Stability passes — several releases focused on correctness over features: output decoding, partial match handling, stderr counting. The database is now something you can actually trust as evidence.
What's coming
Native capture
Right now chron works because the model cooperates — it calls chron's MCP tools to log what it's doing. The gap: a model that doesn't call the tools leaves no trace.
Native capture catches sessions at the client level, from hook events fired by the AI tool itself, without depending on the model. The session is recorded regardless of whether the model decides to log it. The difference between "audit when the AI remembers to" and "audit always."
First measured finding against Claude Code is complete. Next: shipping it and testing across more clients.
Local UI
chron ui — a localhost dashboard to browse, search, and inspect your session history. Filter by date, client, session content. See everything your AI tools have done across all sessions in one view. No server, no cloud, reads the same SQLite file.
Prompt injection detection
Recent incidents — secrets exposed via AI agents, agents executing instructions hidden inside documents they were told to summarize — are the AI equivalent of SQL injection. Untrusted content in the model's context that looks like instructions.
Chron's scan_prompt is the foundation. The evolution: scan agent context before it acts, flag content that looks injected, and log it as a threat signal. Chron's job is to witness and record; blocking belongs to an authorization layer.
The constraint that shapes everything
Chron is local-first by conviction, not by default. 13,300+ downloads and we genuinely don't know how many people use it daily — because we built it not to know. No telemetry, no pings, no session data ever leaves your environment.
If you're using chron, or tried it and found something missing, we'd like to hear from you. Native capture and the UI are being shaped by feedback right now.

Top comments (0)