Your agent drafts.
You comment. It revises.
Marigold gives you and your agent a shared artifact: the agent serves a rich, interactive HTML page to your browser, you pin comments right on the rendered page, and every note flows straight back into the agent’s context. No copy-paste, no screenshots, no “line 341 of the diff”.
Copy this into your coding agent — it installs and wires itself up:
Install Marigold for me using `npm i -g marigold-draft`, then read https://marigold.page/draft/setup.md?ref=site-hero and set yourself up to use it.
Or install manually: npm i -g marigold-draft then marigold-draft agent-setup
A review loop, not a chat scroll
Reviewing an agent’s work in chat means walls of text and re-explaining what you meant. Marigold turns it into the loop you already know from Google Docs.
“Code this.” “Help me decide.” “Draft this up.” “Help me understand how it works.” Any task that ends in an artifact.
Not a wall of prose — a decision gets a weighted matrix, a review gets margin annotations, an explainer gets a diagram. One self-contained page, in your browser.
Pin notes to the exact element, edit text inline, then hit “Send feedback to agent”. No screenshots, no “the third paragraph”.
Your feedback lands structured in the agent’s context. It revises, replies to each comment, and your notes re-attach to the new version — nothing is lost.
One command puts the same page at a link anyone can comment on — teammates join the loop, no account needed. That’s hosted Marigold.
…and the doc arrives fit for the job
Marigold helps your agent use the right user interface for the task, so your review is as easy as possible. No more long markdown essays and flat tables for everything.
produces a mental model — layered explainer, master diagram, prediction prompts
produces verdicts on work — the artifact inline, margin annotations, severity colors, per-section sign-off
produces a selection + rationale — trade-off matrix with weight sliders — re-scores as you drag
produces an arrangement — kanban drag, drag-to-rank, impact/effort canvas — pre-sorted with the agent’s best guess
produces parameter values — sliders with the thing you’re tuning re-rendering live
produces a completed procedure — runbook stepper with prerequisites, persisted progress, branch points
produces an updated picture — “what changed since you last looked” first, exception highlighting
Every mode: answer-first, three reading depths, and a structured exit — your corrections end up machine-readable. marigold-draft principles judge prints the pack; any agent can follow it.
Built for agents. Your data stays on your machine.
Open a draft and the agent’s command doesn’t return until you’ve reviewed. Your notes land straight back in its context — no pasting, no nudging.
Comments live in a small file next to your draft and re-attach to every new version. Restarts, crashes, closed laptops — the thread survives.
Everything runs on your machine. No account, no cloud — your drafts and your data never leave localhost. The CLI reports anonymous usage counts, which you can turn off.
Works with any MCP-capable coding agent:Claude CodeCodexAntigravityopencode
Free. MIT-licensed. Yours.
Everything above is the local tool, and the local tool is free forever — MIT open source, so you can audit it, fork it, and ship it inside your company without a procurement conversation.
Install in 30 seconds
Install Marigold for me using `npm i -g marigold-draft`, then read https://marigold.page/draft/setup.md?ref=site-hero and set yourself up to use it.
Click to select, paste into Claude Code (or any coding agent) — it installs the tool and wires itself up.
$npm i -g marigold-draft+ marigold-draft added in 4s$marigold-draft agent-setup✓ agent skill installed · MCP server registered$
Node 20+. Full setup guide: /draft/setup.md.
When you’re ready
Then share it with humans.
Drafts live on your machine. When one is worth keeping — or a teammate should see it — hosted Marigold takes the same HTML unchanged: marigold-draft share draft.html returns a link anyone can comment on, no account needed. Claim it into a free account for saved versions, sharing by email, and the same agent loop over MCP.