# aiSecurityWarden > A secrets manager and approval layer for AI agents. API keys and passwords are encrypted on the user's own > computer before they're stored, so the service only holds data it can't read. Agents get scoped, revocable > tokens instead of keys, and sensitive secrets and actions (sending email, deleting, merging code) wait for the > person to tap Approve on their phone or the website. Website: https://aisecuritywarden.com/ Web app: https://app.aisecuritywarden.com/ Open-source client (Apache-2.0): https://github.com/SoftDevEng/aiSecurityWarden-client Install the client: `pipx install aisecuritywarden` (https://pypi.org/project/aisecuritywarden/); step-by-step guide for Windows, macOS and Linux, including connecting Claude Desktop, Claude Code and Hermes: https://aisecuritywarden.com/install.html The full guide (projects and secrets, tokens, approvals, action rules, connectors, root commands with sw-sudo, security model, what is stored): https://aisecuritywarden.com/help.html ## What it does - End-to-end encrypted secrets: encryption happens in the client or the browser; the server stores ciphertext, wrapped keys and password-proof hashes only. - Machine tokens for bots and agents: scoped to one project, limited to allowed addresses, with expiry dates, one-time setup codes and instant revocation. - Human approval: a token can require the owner's approval once per session or for every fetch. Requests show a match code and arrive on the website, a passkey prompt, Telegram, ntfy or Discord. - AI agent action approval: through MCP connectors (Gmail and GitHub built in, plus custom connectors and an approval proxy in front of any MCP server), each action is checked against the owner's rules: allow, ask every time, allow for a time window, or deny. A denial can carry a comment telling the agent what to change. - Gmail connector: search, read (including the text of PDF, Word and text attachments), organize, draft, send, reply, forward; attachments the owner approves file by file, and big files as Google Drive links shared view-only with the email's recipients only. - Tool-change protection: if a connected MCP server changes what its tools say, nothing new is offered to the agent until the owner approves the new tool list. - Activity log, two-factor sign-in, passkeys, and signed client updates (`sw-mcp update`). ## Works with - Any agent that speaks MCP (Model Context Protocol); tested with Claude Desktop and Hermes Agent. - Bots and scripts without MCP: `swctl export` / `swctl run` hand secrets to a program at start-up. - Windows, macOS and Linux. ## Plans A permanent free plan (1 user, 3 projects, 50 secrets, 3 machine tokens) with every security feature; paid plans add volume, phone push approvals, longer history, log export and team features. See https://aisecuritywarden.com/#pricing. ## Contact Use the form at https://aisecuritywarden.com/#contact (sales, support, privacy, security reports, investors and partners).