secrets-mcp

An encrypted credential vault for Claude Code (and any MCP client): store and use API keys, passwords, tokens, SSH keys, TLS certificates, and other sensitive material without the values ever entering the model's context.

The model can list what exists, collect new credentials from the user, run commands with credentials injected, call APIs with stored tokens, and render config files — but it never sees a value. Ships as a Claude Code plugin bundling the MCP server, a guard hook, and a skill.

How values stay out of the model

• Input never crosses the chat. vault_set opens a native dialog (hidden input on macOS) where the user types the value; vault_import reads a file straight from disk into the vault. Headless/Linux/Windows: the secrets-vault CLI reads from a hidden TTY prompt or piped stdin — never argv.

• Output is redacted. vault_run and vault_http scrub every stored value from captured output before it is returned — raw, base64, base64url, hex, URL-encoded, and JSON-escaped forms, plus each individual line of multi-line values (so grep/head on a PEM file comes back redacted).

• Disk is guarded. A PreToolUse hook denies the model Read/Edit/Grep/Bash access to the vault directory and to any file produced by vault_render or vault_write.

• Storage is encrypted. AES-256-GCM vault files; the master key lives in the platform credential store — macOS Keychain, Linux libsecret (secret-tool), Windows DPAPI — with a 0600 key file as last-resort fallback. Key material moves over stdin/stdout, never argv.

Related MCP server: API Locker

Threat model — read this

This protects against accidental exposure: secrets pasted into chat, echoed by commands, logged, committed, or read out of config files into the conversation. It is not a sandbox against an adversarial model. A model that deliberately transforms a secret before printing it (e.g. cut -c1-4, character arithmetic, XOR) defeats exact-match redaction — the redaction layer catches common encodings and line-wise extraction, not arbitrary computation. If that is your threat, the mitigation is Claude Code's permission prompts on vault_run commands, not this tool.

Materialized files (vault_render / vault_write outputs, files: temp paths) are plaintext on disk while they exist; temp files are 0600, created inside the guarded directory, and deleted when the command exits.

Entry model

Non-credential types (ip_address, hostname, url, username, email) get a visible input dialog but are stored, redacted, and injected exactly like secrets — useful for infrastructure details that should not appear in transcripts.

MCP tools

Example — the model deploys over SSH without ever holding the key:

vault_run {
  "command": "ssh -i \"$KEY\" -o IdentitiesOnly=yes deploy@prod 'systemctl restart app'",
  "files": { "KEY": "deploy-ssh-key" }
}

CLI

For terminals, SSH sessions, and multi-line values:

$ secrets-vault set github-token --type api_key --scope global
Value for 'github-token' (input hidden): ...
$ secrets-vault import deploy-key --file ~/.ssh/id_deploy --type ssh_private_key
$ pbpaste | secrets-vault set staging-db --type connection_string
$ secrets-vault list
$ secrets-vault rm old-token --scope project

Install

$ npm install && npm run build

As a Claude Code plugin (MCP server + guard hook + skill). Plugins install from marketplaces, and this repo self-hosts as one:

$ claude plugin marketplace add DatanoiseTV/secrets-mcp
$ claude plugin install secrets-mcp@datanoise --scope user

From a local clone (development):

$ claude --plugin-dir /path/to/secrets-mcp        # try it without installing
$ claude plugin marketplace add /path/to/secrets-mcp
$ claude plugin install secrets-mcp@datanoise --scope user

Both paths work out of the box: dist/ is a committed, self-contained esbuild bundle with no runtime dependencies — installing the plugin requires no npm step, only Node >= 20 on PATH.

Or just the MCP server, without the plugin:

$ claude mcp add --scope user secrets -- node /path/to/secrets-mcp/dist/index.js

Permissions: skipping the "allow this tool?" prompts

Claude Code asks before every MCP tool call unless the tool is allowlisted. For the plugin, the tools are named mcp__plugin_secrets-mcp_secrets__<tool> (standalone claude mcp add install: mcp__secrets__<tool>). Add rules to ~/.claude/settings.json (user-wide) or .claude/settings.json (per project), or interactively via the /permissions command.

Recommended allowlist — everything whose consent already happens elsewhere or that cannot expose a value:

{
  "permissions": {
    "allow": [
      "mcp__plugin_secrets-mcp_secrets__vault_list",
      "mcp__plugin_secrets-mcp_secrets__vault_check",
      "mcp__plugin_secrets-mcp_secrets__vault_set",
      "mcp__plugin_secrets-mcp_secrets__vault_import",
      "mcp__plugin_secrets-mcp_secrets__vault_delete",
      "mcp__plugin_secrets-mcp_secrets__vault_render",
      "mcp__plugin_secrets-mcp_secrets__vault_write",
      "mcp__plugin_secrets-mcp_secrets__vault_cleanup"
    ]
  }
}

vault_set, vault_import (with remove_source), and vault_delete open native dialogs — the dialog is the consent, so auto-allowing the tool call loses nothing. vault_list and vault_check return metadata only.

Deliberately not on the list:

• vault_run — executes arbitrary shell with secrets injected. The permission prompt showing you the command is the backstop against a malicious or confused command extracting a value past the redaction layer (see threat model). Keep it prompted.

• vault_http — sends stored credentials to whatever URL is in the call. A prompt showing the destination is what stands between a stored token and an unintended host. Keep it prompted.

If you accept those risks in a trusted workflow, one rule allows the whole server: "mcp__plugin_secrets-mcp_secrets".

Environment

Development

$ npm test          # vitest: crypto round-trips, scoping, redaction, injection, cleanup

Tests run against an isolated SECRETS_MCP_HOME with an env master key; they never touch the real keychain or vault.