MCP Server

Audex ships as an MCP server so Claude Code can request scoped, short-lived credentials natively for AWS, GCP, Azure, or Vault — without Bash wrapping.

---

Setup

Add to your project’s .mcp.json:

{
  "mcpServers": {
    "audex": {
      "command": "/path/to/audex",
      "args": ["mcp"]
    }
  }
}

Set provider-specific environment variables before starting Claude Code:

For AWS:

export AUDEX_ROLE_ARN="arn:aws:iam::123456789012:role/AudexAgentRole"

For GCP:

gcloud auth application-default login
export AUDEX_GCP_SERVICE_ACCOUNT="audex-agent@my-project.iam.gserviceaccount.com"

For Azure:

az login
export AZURE_SUBSCRIPTION_ID="12345678-1234-1234-1234-123456789012"

When Claude Code starts, it connects to the Audex MCP server automatically. Claude then sees three native tools.

---

Tools

audex_run

Run a command with scoped, short-lived credentials for AWS, GCP, Azure, or Vault.

Parameter	Type	Required	Description
allow	string	Yes	Comma-separated IAM/permission actions (format depends on provider)
command	string[]	Yes	Command and arguments to execute
provider	string	No	Cloud provider: “aws”, “gcp”, “azure”, “vault” (default: “aws”)
ttl	string	No	Time-to-live (default: “15m”)
budget	number	No	Budget limit in USD (AWS only)
role_arn	string	No	IAM role ARN (AWS) or service account email (GCP)
profile	string	No	Named policy profile (e.g., “s3-readonly”)
resource	string	No	Resource ARN restriction

Examples:

// AWS: read S3
audex_run({
  provider: "aws",
  allow: "s3:GetObject,s3:ListBucket",
  command: ["aws", "s3", "ls"]
})

// GCP: read Cloud Storage
audex_run({
  provider: "gcp",
  allow: "storage.objects.get,storage.objects.list",
  command: ["gsutil", "ls"]
})

// Azure: list storage accounts
audex_run({
  provider: "azure",
  allow: "Microsoft.Storage/storageAccounts/read",
  command: ["az", "storage", "account", "list"]
})

// Using a profile
audex_run({
  provider: "aws",
  profile: "s3-readonly",
  command: ["aws", "s3", "ls"]
})

audex_sessions

List credential sessions with their status and metadata.

Parameter	Type	Required	Description
status	string	No	Filter by status
limit	number	No	Max results (default: 20)

audex_audit

View the audit trail of session events.

Parameter	Type	Required	Description
session_id	string	No	Filter by session ID
limit	number	No	Max entries (default: 50)

---

How It Works

The MCP server runs as a long-lived process communicating over stdin/stdout using JSON-RPC (the Model Context Protocol). It is built in Rust using the rmcp crate (the official MCP SDK). No Anthropic API key is required.

When Claude calls audex_run:

1. Audex creates a session and issues STS credentials (same as tryaudex run CLI)
2. The command runs as a subprocess with credentials injected
3. stdout and stderr are captured (not inherited, since stdin/stdout is the MCP channel)
4. Results are returned to Claude as structured text
5. Session is logged to the audit trail

---

Differences from CLI

Behavior	CLI (tryaudex run)	MCP (audex_run)
stdio	Inherited (you see output live)	Captured (returned to Claude)
Interaction	User runs it manually	Claude calls it as a tool
Output	Printed to terminal	Returned as MCP tool result
Session tracking	Same	Same
Audit logging	Same	Same

---

Testing the MCP Server

You can test the server manually by sending JSON-RPC messages:

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}' | tryaudex mcp

The server should respond with its capabilities and tool definitions.