sidewalk

Your agent finishes a change and, instead of claiming it works, writes you a walkthrough through an MCP server: cards that open in a Chrome side panel beside the running site. Press Go and the panel takes the tab to the page a card names and outlines what to look at. You answer Pass, Issue or Skip, and the verdict goes back with a screenshot, the page's console tail and whatever you typed. A card can instead be a sequence: one page, a few ordered steps you tick as each works, so an Issue says which step broke. Or a question, when the agent wants a decision rather than a check. Ask runs the other way, and the answer comes back on the card. Every answer is yours for ten seconds before any agent can read it, and Undo takes it back. A card can carry a demo key as dots with a Copy button, kept out of the record. sidewalk is the panel half; walkd, a daemon on your own machine, is the other, and the panel talks to 127.0.0.1 and nothing else. Node 22 or newer: claude mcp add --scope user sidewalk -- npx -y sidewalk-mcp walkd mints a token on first start; the panel asks for it once. A token is not a login: anything that can read your files can read it, so walk with demo keys and test accounts. Writes stay under your own data directory. No sidewalk server, no analytics. MIT, source at github.com/kesensoy/sidewalk.