Switch AI coding agents (Claude Code, Codex, Copilot) mid-task without losing your place.
DevRelay keeps task state, decisions and progress in a .aiw/ folder inside your repo. Agents read and update it
through a local MCP server, and the extension coordinates one-click handoffs between them.
Features
Shared task state: one active task with goal, steps, next action and blockers, visible in the DevRelay activity-bar view.
One-click handoff: DevRelay: Hand off to... opens a preview of the exact context the next agent will receive,
with per-section token costs, a budget meter and checkboxes to trim sections.
Decision log: agents record decisions in .aiw/decisions.md so the next agent doesn't re-litigate them.
Auto-checkpoints: mechanical progress entries (diff stat, test status, open files; no LLM call) on new commits,
finished test runs, 15 minutes of edits and window close.
MCP server: tools such as get_context, start_task, update_task, record_decision, checkpoint and
search_history. Copilot agent mode picks it up automatically; Claude Code and Codex are registered with
DevRelay: Connect agents.
Chat participant: @devrelay answers questions about the current task, searches history (/search) and
reviews uncommitted changes with several models (/review).
Getting started
Open a git repository.
Click the DevRelay icon in the activity bar, then Initialize DevRelay. This creates .aiw/ and adds an
instructions block to AGENTS.md.
Run DevRelay: Connect agents (Claude Code, Codex).
Start task, work with any agent, then Hand off to... when you want to switch.
Requirements
VS Code 1.101 or later
Optional: Claude Code (claude) and/or Codex CLI (codex) for handoffs to those agents
Privacy
Everything stays in your repo and on your machine. Recent conversation turns are off by default; if you enable
them, only plain-text turns are read from local session files, redacted and truncated. API keys for the review
feature are stored in VS Code's secret storage.
Settings
Setting
Default
Meaning
devrelay.anthropicModel
claude-sonnet-5-5
Model used with your Anthropic API key
devrelay.openaiModel
gpt-4.1
Model used with your OpenAI API key
Known limitations
Window-close checkpoints are best effort.
Single-root workspaces only; multi-root uses the first folder.
Re-run Connect agents after updating the extension.
Token counts are estimates (about 4 characters per token).
