How I use Relay with Claude Code every day

I build Relay with Claude Code, and Relay itself is the memory layer for that work. Very meta, but it means I feel every rough edge myself. Here's the actual workflow, not the marketing version.

Setup is one command

npx @onrelay/wizard detects your coding agents (Claude Code, Cursor, Codex, Windsurf and others), installs the MCP server, and signs you in. Takes about a minute. After that your agent has six Relay tools: get the brief, recall context, search sources, save memory, list projects, switch projects.

Session start: get_brief

Every session starts with the agent calling get_brief. It gets back the project objective, active constraints, open tasks, and what happened in recent sessions. No "let me explore the codebase for 10 minutes to figure out what we're doing" — it already knows the decisions behind the code, which a codebase scan can't tell you.

During work: it saves what matters

When we decide something durable ("personal memory is a project with kind=personal", "cron maxDuration caps at 60 on hobby plan"), the agent saves it as a decision or constraint. When something blocks us, it becomes a task. I don't think about this, it just happens as part of the work.

The killer feature: surviving compaction

Long sessions get compacted and agents forget the middle of the conversation. Relay hooks into Claude Code's lifecycle (PreCompact, SessionEnd) and checkpoints context before it's lost. Next session picks up exactly where the last one ended, even across machines.

And it's the same memory my chats use

This is the part that compounds: when I brainstorm a feature in ChatGPT on my phone, the extension captures the decisions. When I sit down at my desk, Claude Code already has them via MCP. One memory, every tool.

If you work with a coding agent daily, try this loop for a week. The difference shows up on day two, when your agent opens with "continuing from yesterday: we decided X, next step is Y" instead of asking what the project is about.