Always-on

Agent Relay's always-on layer captures context from every AI coding agent you use and hands off automatically when one rate-limits. No per-session command. Nothing to remember.

The original CLI is still here — see relay c, relay handoff. The always-on layer adds a second surface: a small background daemon, four adapters, and an HTTPS proxy opt-in. Everything writes to the same on-disk store, and both surfaces read from it.

What it does

End-to-end

terminal
curl -fsSL https://agent-relay.dev/install.sh | sh   # native binary
relay install                                         # wires every adapter
# (reboot — or just `relay daemon start` to skip the wait)

After that: open Cursor at 9am, switch to Claude Code at 11am, run aider in Warp at 2pm. You never type relay. At 3pm Claude Code hits its limit; a notification fires, you click "Resume in Cursor," and the next agent opens with everything pre-loaded.

The four adapters

AdapterWhat it coversDocs
Claude Code hookClaude Code anywhere/adapters/claude-code
PTY wrapper (relay wrap)Every CLI agent — codex, aider, gemini-cli, sgpt, llm/adapters/cli
VS Code extensionCursor, Antigravity, VS Code, Windsurf, Trae, Void/adapters/cursor
Warp MCPWarp Agent Mode/adapters/warp

One hook, one wrapper, one extension, one MCP server — together they cover the modern agentic-coding stack.

How to verify it's working

terminal
relay doctor          # checks every adapter + the daemon
relay daemon tail     # live event stream
relay snapshots       # list captured handoff primers
relay dashboard       # local web UI

See Architecture for the moving pieces and where they live on disk.