agentic-relay

A unified CLI that brings Claude Code, Codex CLI, and Gemini CLI under a single interface -- solving tool fragmentation, enabling multi-layer sub-agent orchestration via MCP, and providing proactive context window monitoring.

Why agentic-relay?

Working with multiple AI coding CLIs means juggling different command syntaxes, configurations, and authentication flows. Nesting sub-agents across tools is not supported natively. And context window exhaustion catches you off guard.

agentic-relay addresses three problems:

Prerequisites

• Node.js 22+
• One or more backend tools installed: Claude Code, Codex CLI, or Gemini CLI
• macOS 13+ / Ubuntu 22.04+ / Debian 12+

Installation

# Install globally
npm install -g @rk0429/agentic-relay

# Initialize project-level config (optional)
relay init

# Verify installation
relay doctor

From Source

git clone https://github.com/RK0429/agentic-relay.git
cd agentic-relay
pnpm install
pnpm build
pnpm link --global

Quick Start

# Start an interactive session with Claude Code
relay claude

# One-shot prompt with Codex CLI
relay codex -p "Refactor this function to use async/await"

# Continue the latest Gemini session
relay gemini -c

# Check which backends are available
relay doctor

Usage

Backend Commands

relay claude|codex|gemini              # Interactive mode
relay claude|codex|gemini -p "query"   # Non-interactive (one-shot)
relay claude|codex|gemini -c           # Continue latest session
relay claude|codex|gemini -r ID        # Resume session by ID
relay claude|codex|gemini --agent NAME # Specify agent
relay claude|codex|gemini --model NAME # Specify model

Management Commands

relay config show               # Show merged config
relay config get <key>          # Get config value
relay config set <key> <value>  # Set config value
relay auth <backend> login      # Authenticate a backend
relay auth <backend> status     # Check auth status
relay sessions                  # List/search sessions
relay update                    # Update backend tools
relay version                   # Show versions
relay doctor                    # Run diagnostics
relay init                      # Initialize .relay/ config

MCP Commands

relay mcp list              # List MCP servers
relay mcp add <name> -- CMD # Add MCP server
relay mcp remove <name>     # Remove MCP server
relay mcp sync              # Sync MCP config to backends
relay mcp serve                              # Start as MCP server (stdio, default)
relay mcp serve --transport http             # Start as MCP server (HTTP)
relay mcp serve --transport http --port 8080 # Custom port

Configuration

agentic-relay uses a three-tier configuration system. Each tier overrides the one above it.

Example configuration:

{
  "defaultBackend": "claude",
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"]
    }
  },
  "hooks": {
    "hooks": [
      {
        "event": "ContextThreshold",
        "command": ["node", "scripts/save-state.js"],
        "timeoutMs": 5000,
        "onError": "warn"
      }
    ]
  },
  "contextMonitor": {
    "enabled": true,
    "thresholdPercent": 75,
    "notifyMethod": "hook"
  }
}

Environment Variables

Security Considerations

• Claude adapter permission bypass: By default, the Claude adapter runs with bypassPermissions mode to enable non-interactive sub-agent execution. This means spawned Claude Code agents can execute tools without user confirmation. To change this behavior, set the RELAY_CLAUDE_PERMISSION_MODE environment variable to default.

MCP Server Mode

agentic-relay can act as an MCP (Model Context Protocol) server, allowing any MCP-capable client to spawn sub-agents across all three backends. This is how nested sub-agent orchestration works -- a parent agent calls relay via MCP, which spawns a child agent on any backend, and that child can call relay again to spawn a grandchild.

Setup

Add relay as an MCP server in your client's configuration. For Claude Code:

{
  "mcpServers": {
    "relay": {
      "command": "relay",
      "args": ["mcp", "serve"]
    }
  }
}

Exposed Tools

Recursion Guard

The MCP server includes a three-layer recursion guard to prevent runaway agent chains:

1. Depth limit -- rejects calls that exceed the maximum nesting depth (default: 5)
2. Per-session call limit -- caps the total number of spawns within a single session (default: 20)
3. Loop detection -- identifies repeated calls with the same backend + prompt hash

Context propagation between parent and child agents uses environment variables: RELAY_TRACE_ID, RELAY_PARENT_SESSION_ID, and RELAY_DEPTH.

Multi-layer Agent Call Flow

sequenceDiagram
    participant User
    participant Parent as Parent Agent<br>(Claude Code)
    participant Relay as agentic-relay<br>(MCP Server)
    participant Child as Child Agent<br>(Codex CLI)
    participant Relay2 as agentic-relay<br>(MCP Server)
    participant Grandchild as Grandchild Agent<br>(Gemini CLI)

    User->>Parent: Task request
    Parent->>Relay: spawn_agent(codex, prompt)
    Note over Relay: RecursionGuard check<br>depth=1, calls=1
    Relay->>Child: Codex SDK thread.run("prompt")<br>RELAY_DEPTH=1
    Child->>Relay2: spawn_agent(gemini, sub-prompt)
    Note over Relay2: RecursionGuard check<br>depth=2, calls=1
    Relay2->>Grandchild: gemini -p "sub-prompt"<br>RELAY_DEPTH=2
    Grandchild-->>Relay2: Result
    Relay2-->>Child: Result
    Child-->>Relay: Result
    Relay-->>Parent: Result
    Parent-->>User: Final answer

Architecture

agentic-relay is organized into six layers.

graph TD
    A["CLI Layer<br><i>citty entry point</i>"] --> B["Command Handlers<br><i>backend, mcp, config, auth,<br>update, sessions, version, doctor, init</i>"]
    B --> C["Core Layer<br><i>SessionManager, ConfigManager,<br>AuthManager, HooksEngine,<br>ContextMonitor, EventBus</i>"]
    B --> D["MCP Server<br><i>RelayMCPServer,<br>spawn_agent, list_sessions,<br>get_context_status, RecursionGuard</i>"]
    C --> E["Backend Adapters<br><i>BaseAdapter, ClaudeAdapter,<br>CodexAdapter, GeminiAdapter,<br>FlagMapper</i>"]
    C --> F["Infrastructure<br><i>ProcessManager, Logger</i>"]
    D --> E
    D --> F

Backend Adapters

Each backend CLI has different flags, output formats, and session handling. The adapter layer normalizes these differences behind a common BackendAdapter interface. Adding a new backend means implementing this interface and registering it with the AdapterRegistry.

Non-interactive execution uses official SDKs where available:

Hooks Engine

An event-driven hook system that executes external commands via stdin/stdout JSON pipes.

Context Monitor

Monitors context window usage and fires ContextThreshold events when usage exceeds the configured threshold. Notification methods: stderr (direct output) or hook (triggers the Hooks Engine).

Project Structure

src/
  bin/relay.ts                  # Entry point
  commands/                     # Command handlers
    backend.ts                  #   relay claude/codex/gemini
    mcp.ts                      #   relay mcp (list/add/remove/sync/serve)
    config.ts                   #   relay config
    auth.ts                     #   relay auth
    update.ts                   #   relay update
    sessions.ts                 #   relay sessions
    version.ts                  #   relay version
    doctor.ts                   #   relay doctor
    init.ts                     #   relay init
  core/                         # Core modules
    session-manager.ts
    config-manager.ts
    auth-manager.ts
    hooks-engine.ts
    context-monitor.ts
    event-bus.ts
  adapters/                     # Backend adapters
    base-adapter.ts
    claude-adapter.ts
    codex-adapter.ts
    gemini-adapter.ts
    adapter-registry.ts
    flag-mapper.ts
    install-guides.ts
  mcp-server/                   # MCP server mode
    server.ts
    recursion-guard.ts
    tools/
      spawn-agent.ts
      list-sessions.ts
      get-context-status.ts
  infrastructure/               # Infrastructure
    process-manager.ts
    logger.ts
  types/                        # Type definitions
  schemas/                      # Zod validation schemas

Tech Stack

• Runtime: Node.js 22+
• Language: TypeScript
• Package manager: pnpm
• CLI framework: citty
• Bundler: tsup (esbuild-based)
• Backend SDKs: @anthropic-ai/claude-agent-sdk, @openai/codex-sdk
• MCP: @modelcontextprotocol/sdk
• Process management: execa (interactive modes, Gemini CLI)
• Validation: zod
• Logging: consola
• Testing: vitest (771 tests across 35 files)
• Coverage: @vitest/coverage-v8

License

Apache-2.0