Two things I ve settled on after a lot of agent work: 1. For most outputs, reviewing a fit-for-purpose rich HTML doc beats reviewing markdown (diagrams, charts, dynamic elements). 2. Feedback works best as in-line comments on the output, rather than re-describing what your feedback applies to.Marigold wires both into one loop. Your agent drafts output as a local HTML file. The file opens in your browser with a comment layer on top. You click an element, type some feedback, and hit send. Your CLI agent gets the feedback, anchored to the elements you clicked. When it updates the file, the tab live-reloads and your comments re-anchor to the new version.The net effect: collaborating with your agent is a few clicks instead of a few paragraphs, and everyone moves faster.Setup is one command: `npm i -g marigold-draft`, then `marigold-draft agent-setup` wires up any CLI assistants are on your machine — Claude Code, Codex, opencode. Or you can find a blurb to paste into your agent on the website.---Some details I cared about while building it:- Everything runs on localhost when you re workign with your agent. No account, no server, nothing leaves your machine.- MIT licensed free forever: https://github.com/immuneeb/marigold-collab- Comments anchor by a composite strategy, so they survive the agent rewriting the page.- There s an optional hosted version for sharing a local Marigold with other humans and their agents.This has helped me move faster and enjoy working with my agents more, hope you find it helpful too. Happy to answer questions.