Spacesheep: What Claude Artifacts Could Have Been

Claude Artifacts could have been this. Spacesheep starts in the same place, an agent makes a page and you share it, and then keeps going. Any agent, on any model, from any tool, publishes to a real URL. Every deploy is a named version. People react and comment on single blocks, and the agent answers over MCP. Add public profiles, friends, teams, live pages and proper privacy controls, and it’s starting to feel like a social platform for agent-made pages.

Key Takeaways

• Model-agnostic: Spacesheep’s own agent runs on Claude, ChatGPT, Gemini, Grok, Kimi, DeepSeek or Nemotron, with your key
• Harness-agnostic: publish from Claude Code, Codex, Cursor, Gemini CLI, any MCP client or the MIT-licensed CLI, or just message the agent on Telegram, Slack or WhatsApp
• Social by default: profiles with listed pages, public search, reactions, block comments, friends, teams and groups
• Privacy you can tune: four visibility tiers, view or edit roles, a default audience for new pages, URL-locked secrets and scoped API keys
• Shared Claude artifacts need a Claude account to open, except legacy chat artifacts (Claude Help Center, 2026)

What is Spacesheep, and why are we writing about it?

Spacesheep calls itself “a shared workspace for people and their AI agents” (Spacesheep, 2026). An agent drops an HTML file or a folder at spacesheep.dev/@handle/slug. People read it, comment, share it and keep editing it, “with every version kept.” Agents connect through one remote MCP server or the spacesheep CLI. Humans just need a browser. It’s built by Michael Makarov and is part of YC’s current batch.

We’ve been using Spacesheep for real Startrise work, and it holds up. We put together a quick Startrise look and published five Labs pages with it. Three are the ones we feature on our profile, so start there:

• Benchmark Explorer: the interactive one. Re-weight our benchmark and watch 15 models re-rank.
• How we run LLM studies: our method on a single page.
• Agent approvals, on one page: the short version of our approval guide.

The other two are our August model-release results from the frontend LLM benchmark and a sneak peek at the Local Code study.

Why does it feel more like a platform than a share button?

Because the page is where things start, not where they end. Tick “show it on my profile” and a public page is listed on your profile, in the sitemap and in search engines (Spacesheep docs, 2026). Anyone can search every discoverable public space by what’s inside it, not just the title. Readers leave emoji reactions and comments on single blocks, and the owner sees who read it and how far they got.

Inviting someone from the web app also makes you friends. Teams get a shared home, and can let anyone with a company email join on their own. Saved groups let you share with a whole list of people at once, and config groups hand a bundle of API keys to the people in them, so they can run the same agent setup.

Then there’s the live stuff. Ask for a live link and you get a “working on it” page in about a second, and live_push updates every open tab in about 100 ms without a redeploy. Streams pipe live numbers from a server or a sensor into a page. Worker apps add server code and secrets, and pages can have built-in storage and realtime rooms. Cloud boxes are Linux containers with Claude Code inside that sleep when idle. Any page prints cleanly or exports to PDF, and on Pro you can copy a space you’re allowed to copy.

What do Claude Artifacts do well today?

Credit where it’s due. Anthropic describes artifacts as “anything Claude makes for you that you’d put in front of someone” (Claude Help Center, 2026). Think a design, a deck, a doc, a dashboard or a small tool, right next to the chat. They’re on every plan, including Free, and in Claude Code. And they’ve grown up a lot.

On paid plans you get templates (Claude Design, Slides and Docs, in beta) with direct editing on the page. Exports go to Word, PDF, Markdown, Google Docs, PowerPoint or plain HTML, depending on the template. Artifacts can store data (20 MB, text only) and plug into apps like Asana, Google Calendar and Slack. They can even call Claude, and that usage “counts against each person’s own plan limits rather than yours” (Claude Help Center, 2026). So a shared mini-app costs you nothing.

Sharing’s better too: Can view, Commenter and Can edit access, email invites in beta, and admin controls on Team and Enterprise plans. In Claude Code, an artifact is “a live interactive page at a private URL,” and “the page updates in place as your session continues” (Claude Help Center, 2026).

If your whole team lives in Claude, that’s a solid default.

Spacesheep vs Claude Artifacts, side by side

The big differences are which models and tools can publish, what happens around the page, and who can read it. Artifacts come from Claude’s apps and Claude Code, and every viewer needs a Claude account apart from legacy chat artifacts (Claude Help Center, 2026). Spacesheep takes pages from any MCP client, runs its own agent on seven model providers, and has a no-sign-in tier for static pages (Spacesheep docs, 2026).

Three things are why we reached for Spacesheep.

Why does working from any agent and any model matter?

Spacesheep documents setup for Claude Code, Codex, Cursor, Windsurf and Gemini CLI, plus the Claude.ai, ChatGPT and Grok connectors and an xAI API config (Spacesheep docs, 2026). Then there’s Spacesheep’s own agent, the one you message on Telegram, Slack or WhatsApp. You pick what it runs on: Claude, ChatGPT, Gemini, Grok, Kimi, DeepSeek or Nemotron, on your own key. That matters to us because Startrise work never happens in one tool. Our benchmark runs a dozen models through one pipeline, and our agents live wherever the job is. A publishing layer tied to one vendor means copying pages between worlds.

The pages in this post were published by an agent that isn’t Claude, from a box with a shell, through the same MCP server Claude Code would use. Same URLs, same version history, same comment threads.

Why do the privacy settings matter?

Because “anyone with the link” shouldn’t be the only option. Every space sits on one of four tiers. Private means only people you invite, and only you can invite. Members means your invitees can invite others. Signed-in means anyone with the link once they sign in. Public means anyone with the link, no sign-in, for static pages. You share with emails, @usernames or saved groups, as view (read and comment) or edit. New pages start private, and you can change that default.

The same care shows up under the hood. Secrets for Worker apps can be locked to a list of allowed URLs, so a key can’t be sent anywhere else. API keys come scoped, too: full, sessions-only for a machine’s coding-session hooks, or streams-only for a box that pushes live values. If a sessions-only key leaks, it can’t do anything but report sessions.

Why do versions and block comments matter?

Every deploy gets a version_name, and the old files stick around. A base_sha guard rejects a deploy that would overwrite someone else’s newer change (Spacesheep docs, 2026). Pages carry data-ss-id anchors, so you can pin a comment to one chart or paragraph. The agent picks it up with list_comments and replies with add_comment.

That’s the loop we care about in human-in-the-loop agent design: you review in place, the agent acts on it, and the history shows who changed what. The Anthropic pages we read cover access levels and comments for artifacts, but not a named version history, so we’ve left that row at “branch the chat.”

What did we learn publishing five pages with it?

Publishing was quick, and it stayed quick. We uploaded our shared assets once. After that, each new page was one index.html reusing the same uploads, which Spacesheep keeps for 24 hours. New spaces started private, and redeploys never touched who could see them. The docs promise exactly that: “a redeploy never un-shares anything” (Spacesheep docs, 2026).

The screenshot tool shows the real page at phone width (390 px) and on desktop. It caught a cramped four-column table on phones, and we fixed it before calling anything done.

The explorer shows how far a static page can go. It’s a single HTML file with the scores baked in as JSON and re-ranked right in your browser, so it can sit on the public, no-sign-in tier. Each leaderboard row has a data-ss-thing name, so a comment sticks to its model even when the table re-sorts. One surprise: spaces load inside a frame that passes the query string through but not the # fragment, so share links had to use ?w=… instead.

A couple of rough edges, to be fair. The deploy linter flagged our CSS variables as “never declared” because they lived in a linked stylesheet. A two-line inline :root fixed it. And design_review, the vision-model critique, needs an LLM key saved in your Spacesheep secrets. We hadn’t added one, so we checked the screenshots by eye.

Nothing dealbreaking. It’s the kind of thing you send through submit_feedback, which, fittingly, is also an MCP tool.

When should you use Claude Artifacts instead?

If you and everyone reading the page already live in Claude, Artifacts is the simpler pick. You get templates with direct editing, exports to PowerPoint, Word and Google Docs, app connectors, and Claude-powered mini-apps that bill each user’s own plan (Claude Help Center, 2026). That’s a lot for zero setup.

Go with Spacesheep when pages come from more than one agent or model, when readers shouldn’t need a Claude account, when you want finer control over who sees what, or when the page is a living thing: versioned, commented block by block, updated live, and exported to PDF on demand. If your agents publish their own reports, that’s also where approval tiers come in. Publishing privately can be auto-approved. Opening a page up to the world deserves a human tap.

Where Spacesheep gets it right

Building something agents use well is harder than building something people use well. Agents read every word of your tool descriptions, then do exactly what they say. Spacesheep’s are the best we’ve worked with. They bake in mobile-first layout, readable contrast, “never invent numbers” and private-by-default sharing, right where the agent will read them. You can see it in the pages.

Give it a spin at spacesheep.dev. And if you want agents that publish, report and ask before they act, that’s what we build at Startrise.