Claude Code

Claude Code is Anthropic’s CLI for AI-assisted coding. Use varlock to inject ANTHROPIC_API_KEY at runtime so the key never sits in a plain text .env or shell history.

For install, schema setup, and shared patterns, see the AI Tools overview.

Environment variable: ANTHROPIC_API_KEY (see supported env variables). Use the API key, not CLAUDE_CODE_OAUTH_TOKEN: that (from claude setup-token) only works for headless claude -p requests, not the interactive TUI.

In a project

Add to .env.schema:

# @sensitive @required
ANTHROPIC_API_KEY=op(op://api-local/anthropic/api-key)

Tip

You can use any secrets provider you want, but we like 1Password since it uses biometric authentication for local access. All the examples in this guide use the 1Password plugin (op()) to show more real world examples. See that page for plugin setup and auth options.

varlock run -- claude

From any directory

Personal schema at ~/.env.claude:

~/.env.claude

# @sensitive @required
ANTHROPIC_API_KEY=op(op://api-local/anthropic/api-key)

varlock run -p ~/.env.claude -- claude

Shell alias (optional)

alias vclaude='varlock run -p ~/.env.claude --no-redact-stdout -- claude'

--no-redact-stdout keeps Claude’s own terminal output unredacted while secrets stay out of your shell history.

MCP server secrets

MCP server configs (.mcp.json, ~/.claude.json) support ${VAR} expansion, but the variables are read from Claude Code’s own process environment. That works if you launch claude via varlock run, but not for the desktop app, which is not launched from your shell. Instead, let varlock resolve secrets at the point where Claude Code uses them.

Remote servers: headersHelper

For http (and ws) servers, Claude Code’s headersHelper runs a command that prints a JSON object of headers. It runs when Claude Code connects to the server (once per session that uses it, not per request), and re-runs automatically if a call returns 401 or 403.

Use varlock printenv --template to emit the headers JSON:

.mcp.json

{
  "mcpServers": {
    "my-api": {
      "type": "http",
      "url": "https://mcp.example.com",
      "headersHelper": "varlock printenv --template '{\"Authorization\": \"Bearer {{MY_MCP_TOKEN}}\"}' --escape json"
    }
  }
}

The helper runs from the session’s working directory, so varlock picks up the project’s .env.schema automatically. For user-scope servers, or to be independent of the working directory, add -p /absolute/path/to/project (or -p ~/.env.mcp for a personal schema file). For the desktop app, also use an absolute path to the varlock binary. The standalone binary is the safest choice there: the npm-installed CLI needs node on PATH, which the desktop app may not have.

Claude Code gives the helper 10 seconds to run. A warm varlock cache resolves well within that; a cold resolve that prompts for biometric auth may not, so keep caching enabled for schemas used this way.

Local stdio servers: wrap the command

For stdio servers, wrap the server command with varlock run in the server entry itself. This works no matter how Claude Code was launched, since each MCP server is a child process it spawns:

.mcp.json

{
  "mcpServers": {
    "my-local-server": {
      "command": "varlock",
      "args": ["run", "-p", "/absolute/path/to/project", "--filter", "MY_SERVER_*", "--inject", "vars", "--", "npx", "-y", "some-mcp-server"]
    }
  }
}

• --filter injects only the vars that server needs, so each server gets least-privilege access
• --inject vars keeps the __VARLOCK_ENV blob out of a third-party server’s environment
• The server’s stdout is a pipe, so redaction applies to sensitive values in its output; add --no-redact-stdout if that interferes with your server