Agents can check what exists without seeing values. Trusted commands get scoped env injection, rate limits, and audit trails. One command replaces .env sprawl.
secrets have been found in public MCP config files this year.
Source: GitGuardian. Agent workflows made plaintext .env sprawl more dangerous. AgentSecrets adds scoped access, not copy-paste.
Cursor, Claude Code, and custom agents work directly inside your checkout. If the file is there, it can leak through prompts, logs, traces, screenshots, or a poisoned doc. A plaintext .env turns one task into the whole vault.
If an agent can read the file, a malicious page or doc can ask it to print the values.
Live credentials can end up in prompts, terminal output, recordings, and support screenshots.
Every checkout, branch, and teammate becomes another place to sync, rotate, and clean up keys.
has_secretReturns true / false so agents can branch without seeing values.list_secretsNames only. Useful context without credentials.run_commandOpt-in only. Requires read scope, explicit enablement, and an allowlist.Install once. From any repo, link the vault, import the best env file AgentSecrets finds, and run trusted commands through ax.
# 1 · installnpm install -g @agentsecrets/ax# 2 · link the repo and import detected envax init✓ 7 secrets imported · .env ready to remove# 3 · run trusted code with scoped valuesax run -- ./agent.py→ secrets injected · run audited# 4 · MCP for Cursor / Claudenpx @agentsecrets/mcp
Set up ax from the checkout. Simple pricing: $8 per seat.