Member-only story
The Key Is in the Clipboard
A Vault Your Coding Agent Can Use but Never Read
Coding agents need credentials constantly. A NAS password so ssh can run a backup, a Hugging Face token to upload a model, an API key so curl can talk to an email service. The agent cannot do the job without the secret, and the obvious way to hand it over is the worst one: you copy the value and say “the key is in the clipboard”, or you paste it straight into the chat.
From that moment the secret is part of the conversation. It sits in the model’s context, gets re-read on every turn, survives into compaction summaries, and lands in the session transcript on disk. On my machine those transcripts live under ~/.claude/projects as JSONL files, and nothing about them is designed to keep a password safe. The secret did its job once and now lives forever in a log.
I did not set out to build a secrets manager. I set out to find out where my time and tokens were going, and the secrets problem showed up in the data.
Mining six hundred conversations
Every Claude Code session leaves a transcript behind, and daily use adds up: I had 661 transcripts across 46 projects, with roughly eighteen thousand tool calls between them. That is a record of every instruction I repeated, every tool that failed, and every…
