These tools check the key against a stored hash, so without a limit they could be used to guess it online. The equivalent web and CLI endpoints already had this limit.

When a call is throttled, the tool tells the agent how long to wait and not to retry sooner, so a model can't keep retrying the same rejected key.

This release also derives the User-Agent sent to the ApiVault API from the package version, replacing a hardcoded 1.0.3 that had gone stale.

Learn more in the MCP tools reference.

The choice is stored in the connection's OAuth token, so everything that agent reads or writes stays in that project. You don't need to change your MCP server or client configuration, and existing tokens without a project binding keep working.

Learn more about project-scoped connections.

Learn more in the MCP documentation.

{
  "mcpServers": {
    "apivault": {
      "url": "https://mcp.apivault.tech/mcp"
    }
  }
}

Connector templates and setup commands for Cursor, Claude Desktop, Windsurf, and VS Code now use this URL. RFC 9728 resource and authorization server declarations were also verified across all metadata endpoints.

The security policy contact is now support@apivault.tech.

Learn more in the MCP client setup guide.

Encryption and decryption run on the ApiVault backend. The MCP proxy forwards custom vault keys with requests that need them and does not persist them.

The server hosts RFC 9728 protected resource metadata, designating ApiVault as the OAuth 2.1 authorization server. Self-hosted deployments can specify custom endpoints via MCP_RESOURCE_URL.

Learn more in the MCP documentation.

Security enhancements

• Client registration: Dynamic Client Registration (RFC 7591) validates redirect URIs and lengths. It permits https:, loopback http:, and supported native app schemes, and rejects javascript:, data:, and fragment identifiers.
• PKCE validation: S256 code challenge verification now uses constant-time string comparison.
• Response headers: OAuth token endpoints send Cache-Control: no-store and Pragma: no-cache. The /mcp, /oauth/token, /oauth/register, and metadata endpoints now handle CORS preflight requests.

Learn more about MCP scopes and tools.

You can now connect AI coding agents such as Cursor, Claude Desktop, and Windsurf to ApiVault through a remote Model Context Protocol (MCP) server.

{
  "mcpServers": {
    "apivault": {
      "url": "https://mcp.apivault.tech/mcp"
    }
  }
}

Agents authorize through OAuth with PKCE. You approve the keys:read, keys:write, and keys:reveal scopes the client requests.

Available MCP tools

The server provides six tools:

• list_keys: queries key names, environments, and services without decrypting secrets.
• get_key: retrieves masked key records with metadata.
• reveal_key: retrieves the decrypted value of a key.
• add_key, update_key, and delete_key: manage keys directly from your agent session.

A new Sessions & Devices settings page lets you audit and revoke active CLI and MCP tokens at any time.

Learn more in the MCP documentation.