Ravi

Agent Deck

What Is Agent Deck?

Agent Deck is a TUI-based session manager that gives you a single terminal interface for all your AI coding agents — Claude Code, Gemini, OpenCode, Codex, and more. Think tmux with deep agent awareness built in.

When you’re running 10+ agent sessions across multiple projects, you need visibility and control. Agent Deck provides real-time status (running/waiting/idle/error) across all sessions, instant keystroke switching, fuzzy search, and session forking — each fork inheriting the full conversation history.

Key features:

  • MCP Manager — toggle MCP servers per session or globally; 85–90% memory reduction via Unix socket pooling across sessions
  • Skills Manager — pool-based Claude skills attach/detach per project
  • Git worktrees — isolated working directories per session with merge/cleanup lifecycle
  • Docker sandboxing — run agents in containers with project directories bind-mounted; host tool auth shared automatically
  • Conductors — persistent Claude Code sessions that orchestrate all your other sessions: watch for agents needing help, auto-respond when confident, escalate otherwise, connect to Telegram/Slack for remote control

Where Ravi Fits

MCP Pool Identity Layer

Agent Deck already supports shared MCP servers via Unix socket pooling. Ravi as an MCP provider drops directly in — one Ravi MCP process serving all agent sessions with email, phone, TOTP, and credential lookup. No per-session secret duplication.

The pool architecture is designed exactly for shared infrastructure, and identity is the most shared thing across agents. Every session gets access to the same Ravi identity without separate connections or credential copies.

Sandboxed Credential Injection

Agent Deck’s Docker sandbox shares credentials by extracting Keychain credentials on macOS — brittle and platform-specific. Ravi’s vault is the cleaner answer: mount credentials into containers via the MCP pool. This works cross-platform (Linux CI, remote machines) without Keychain gymnastics.

Conductor Notifications

Conductors connect to Telegram/Slack for remote control. Ravi’s phone and SMS layer is the identity primitive those integrations need — a conductor that needs to SMS a human when something breaks, or verify a service signup during an autonomous run, routes through Ravi rather than borrowing the developer’s personal number.

Setup Guide

1. Install Agent Deck

# Clone and install
git clone https://github.com/asheshgoplani/agent-deck
cd agent-deck
npm install && npm run build

2. Configure Ravi as a Shared MCP Server

Add Ravi to Agent Deck’s MCP pool in ~/.agent-deck/config.toml:

[[mcp_servers]]
name = "ravi"
command = "npx"
args = ["-y", "@ravi/mcp-server"]
scope = "global"          # available to all sessions
pool = true               # share via Unix socket

3. Set Your Ravi Credentials

# Store your Ravi API key so the MCP server can authenticate
export RAVI_API_KEY=your_api_key_here

Or add it to Agent Deck’s environment config so all sessions inherit it.

4. Use Ravi Tools in Any Session

Once the MCP pool is running, any agent session can call:

  • ravi_get_info — get the provisioned email and phone number
  • ravi_inbox_email — check for verification emails
  • ravi_passwords_get — retrieve stored credentials
  • ravi_secrets_get — fetch API keys and tokens

5. Conductor Integration

For conductors that need to notify humans or handle auth walls during autonomous runs, the Ravi MCP tools are available in conductor sessions just like regular agent sessions. A conductor watching for stuck agents can trigger an SMS via ravi_sms_send when human judgment is needed.

Why This Combination Works

Agent Deck solved “share infrastructure across sessions” with MCP pooling. Ravi is the identity infrastructure worth sharing. Together, they give each agent session access to a consistent, accountable identity — without any session needing to manage its own secrets or borrow the developer’s personal credentials.

This is especially valuable in Docker-sandboxed environments where Keychain access is unavailable and agents need a clean, portable way to authenticate against external services.

Resources