AI Agents & MCP

Build secure agent workflows with Envault using the Model Context Protocol (MCP) and Human-in-the-Loop (HITL) approvals.

Envault provides first-class support for AI workflows through our seamless Model Context Protocol (MCP) server. Use your favorite AI coding assistants (like Claude Desktop, Cursor, and RooCode) to fetch, diff, and propose changes to your environment variables without ever leaving your editor.

Because AI agents can hallucinate or overwrite critical production secrets, Envault enforces a strict Human-In-The-Loop (HITL) Approval Model. Agents can propose changes, but a human must explicitly approve them via the CLI or Dashboard before they take effect.

1. Setting up the MCP Server

The absolute fastest way to give your AI assistants access to your Envault environment context is by using the CLI's automated installer.

There are no manual JSON files to edit and no separate tokens to generate. The MCP Server runs as a local shadow daemon and inherits your terminal's secure OAuth session.

Install the MCP Server

To map the MCP Server strictly to your current VS Code workspace, run:

envault mcp install --local

Alternatively, to install the server globally (so every Cursor window or Claude Desktop instance inherently has access to Envault without needing a .vscode/mcp.json file inside the repo), run:

envault mcp install

Restart your Client

After the script finishes injecting the configurations, fully restart your AI application (or reload the VS Code window) to initialize the Envault MCP Server.

When you update the @dinanathdash/envault-mcp-server integration in the future, just run:

envault mcp update

2. Manual Configuration (Standalone MCP Tokens)

For advanced users running isolated IDE instances (like a cloud-hosted DevContainer) where the standard Envault CLI OAuth session (envault login) is unavailable, or if you prefer to edit configurations manually, you can generate a standalone "Web Token".

Generate an MCP Token

Click Generate Token. Copy this securely, as it will only be shown once.

You can only have one active MCP token at a time. Generating a new one instantly revokes the previous token.

Configure your AI Client

Install the server into your AI client of choice.

Edit your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

claude_desktop_config.json

{
  "mcpServers": {
    "envault": {
      "command": "npx",
      "args": ["-y", "@dinanathdash/envault-mcp-server@latest"],
      "env": {
        "ENVAULT_TOKEN": "<YOUR_ENVAULT_TOKEN>",
        "ENVAULT_BASE_URL": "https://www.envault.tech"
      }
    }
  }
}

Open Settings -> Features -> MCP and add a new MCP Server:

Name: envault
Type: command
Command: npx -y @dinanathdash/envault-mcp-server@latest

Then add the exact Environment Variables:

ENVAULT_TOKEN = <YOUR_ENVAULT_TOKEN>
ENVAULT_BASE_URL = https://www.envault.tech

Edit your workspace or global .vscode/mcp.json:

mcp.json

{
  "servers": {
    "envault": {
      "command": "npx",
      "args": ["-y", "@dinanathdash/envault-mcp-server@latest"],
      "env": {
        "ENVAULT_TOKEN": "<YOUR_ENVAULT_TOKEN>",
        "ENVAULT_BASE_URL": "https://www.envault.tech"
      }
    }
  }
}

Windows Users: If npx fails to start within GUI apps (like Claude Desktop), change the command from "npx" to "npx.cmd".

Restart your Client

After modifying the configuration, fully restart your AI application (or reload the VS Code window) to initialize the Envault MCP Server.

3. How the HITL Approval Flow Works

When you ask an AI Agent (like Claude) to "add a STRIPE_KEY environment variable to my production Envault project", the following secure sequence occurs:

Agent Request: The agent uses its MCP envault_push tool to request the mutation.
Envault Intercepts: Instead of updating the database, Envault creates a pending_approvals record and returns a 202 Accepted response.
Agent Pauses: The agent receives an approval_id and an approval_url. It will pause execution or inform you that an approval is required.
Human Verification: You click the approval_url to view the requested changes in the Envault Web Dashboard, or approve it via your terminal:
```
envault approve <approval_id>
```
Execution: Once approved, the mutation is securely executed, and the agent proceeds with its task.

All agent requests, approvals, and rejections are permanently recorded in your Project Audit Logs to ensure total compliance.

4. MCP Tools Reference

Once connected, your AI Agent gains access to the following tools:

Read-Only Tools: envault_status, envault_context, envault_diff, envault_pull
Mutation Tools (Requires HITL): envault_push, envault_deploy, envault_set_local_key, envault_remove_local_key
Execution Tools: envault_run (Injects secrets dynamically into a local command)

Security Kill Switches

If an agent behaves unexpectedly or an MCP token is compromised, you do not need to delete the project.

Envault provides instant Kill Switches:

User Level: Navigate to Settings -> Security and disable "Agent Access". This instantly stops all agents using your tokens across all projects.
Project Level: Workspace Owners can navigate to Project Settings -> Access Control to disable "Agent Workflows", immediately protecting that specific project from all automated AI modifications.

AI Agents & MCP

On this page