Open Artifact gives coding agents a local editor for decks and docs
The Apache-2.0 TypeScript project turns agent output into editable local files, with eight artifact types and a comment loop back to the agent.
Open Artifact for coding agents is an open source project that gives an agent a local editor for the slides, designs, docs and dashboards it produces. JimLiu/open-artifact had 304 stars and 44 forks on 2026-11-22, five days after it was created on 2026-11-17. It is written mainly in TypeScript and released under Apache-2.0.
The idea is narrow and practical. Instead of an agent handing back a file you have to open somewhere else, it writes a folder of plain files that a local page renders as a working editor. You fix a title, move a sticky note, leave a comment, export the result, then send those comments back into the agent's session.
Eight artifact types
The project's README lists eight. Slides make a 1920x1080 deck with a slide rail, overview, speaker notes, builds, transitions and presenter mode, exported to PDF, PPTX, HTML or PNG. Design is an infinite canvas of live .dc.html artboards for screens, components, click-through prototypes and notes, exported to PNG, JPEG, PDF or HTML. Design System is a brand book of tokens, colours with contrast checks, type specimens and live component previews that designs and decks can install; no export formats are listed for it. Whiteboard is an infinite board of stickies, shapes, arrows, ink, images, tables, embeds, live cursors and a timer, exported to PNG.
Sheets is a spreadsheet with several sheets of cells, a formula engine, number formats, merged cells, filtered tables, dropdowns, checkboxes, conditional formats, frozen and grouped rows, charts from cells, CSV import and cell history, exported to XLSX or CSV. Data builds dashboards from CSV, JSON or text uploads or from live MCP connector queries, drawn by a plain HTML/CSS/d3 page, with a Sources view showing where each number came from; it exports to PDF, HTML or JSON. Docs are tabbed documents with suggested edits, comments and version history per tab, importing Markdown, Word and PDF and exporting Markdown, Word, PDF or HTML. Motion makes short animated films from HTML/CSS/JS scenes with a generated score, captions and voiceover, and comments pinned to a moment or a spot in the frame; exports are MP4, GIF, HTML or ZIP.
How the loop runs
You prompt the agent with /open-artifact and a request. The agent writes a folder such as project/deck.json plus slides/*.html. A local daemon serves that folder at http://127.0.0.1: /a/, where the editor opens. When you press "Send to agent", your comments arrive back in the agent's session. Everything runs on your machine, with no account, no upload and no specific model or vendor. The project was formerly called baoyu-design, and the old URL redirects.
Installation needs Node.js 22.18 or later. You download open-artifact-plugins.zip from the Assets section of a GitHub Release and extract it. The source documents plugin steps for Claude Code, Codex and ChatGPT desktop, Cursor, GitHub Copilot CLI, Copilot in VS Code, Antigravity, ChatGPT on the web through a prebuilt connector over HTTPS, and DeepSeek Harness through the dsh-open-artifact npm package. The plugin adds the Artifact, ArtifactData and ArtifactComments MCP tools plus hooks; apps that display MCP Apps also get a ShowArtifact tool that puts the live page in the conversation. For other agents, you copy the codex/skills/open-artifact/ folder into the agent's skill directory, and each type folder can also run on its own through a CLI.
Limits and open questions
A repository five days old with a few hundred stars is early, and the source gives no roadmap, support commitment or release version. It also does not explain what "shared" means for whiteboards and sheets when everything runs locally with no account and no upload, or how a browser-based ChatGPT connector reaches a daemon on your own machine. Large decks, long films and heavy spreadsheets have no stated size limits.
What this suggests for people building apps
This is inference, not something the source states: if artifacts are just folders of files with documented formats, agents can read and rewrite them without a vendor in the middle, which likely makes reviewing and correcting generated output part of the build loop rather than a separate step.
Draftship is an AI app builder where you describe an app and it writes, builds and hosts it, and paid plans add a choice of AI model. If you are weighing which model or agent does the generating, the guide to the AI chat and choosing a model covers that decision. Draftship is not affiliated with this project.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.