Open-source MCP server lets agents use secrets they can never see
JuanixVentures · reddit · 2026-09-26
A developer open-sourced an MCP server (MCPBytes/mcpbytes-vault) that fixes a common leak in agent workflows: when an agent generates a webhook signing key, session secret, or API token, it usually comes back as plaintext in the conversation — landing in chat history, logs, traces, screenshots, or tickets.
The pattern: the agent can use a secret reference but never receives the secret itself.
- The agent requests a secret by label + size
- The secret is generated locally on the user's machine
- Stored in Keychain / Credential Manager / Secret Service / a private file
- The agent only receives a reference like vault:...
- Revealing or deleting the secret requires an interactive terminal
The author is explicit about the threat model: this is not magic isolation — an agent with unrestricted user-account access still needs sandboxing or a separate user. They're soliciting feedback from MCP/agent workflow builders on whether this belongs in tooling.
More from coding & agent
- Anthropic engineer: Claude Tag writes >50% of my PRs every day — bcherny · 2026-09-26
- Cline Desktop adds SSH support: agent runs on remote machines, UI stays local — cpaik · 2026-09-26
- Nautilo: open-source multi-user platform where humans and AI agents share one terminal session — Dan_Jeffries1 · 2026-09-26
- Dev builds 8 Astra skills automating the full creator-marketing pipeline, end to end — alexgoughcooper · 2026-09-26
- Reddit Asks: What's the Worst Thing Your Coding Agent Has Actually Done? — radim11 · 2026-09-26
- One-Day Game Built With Opus 5.5 Stirs Skepticism, Author Promises to Open-Source — tadanada · 2026-09-26