Anthropic ships Claude Code Artifacts as a deliberately stateless canvas

Anthropic shipped Claude Code Artifacts to its Team and Enterprise plans, turning a CLI agent session into a live, shareable HTML workspace that updates in real time as the underlying work progresses. The product arrives more than two weeks after OpenAI's Codex update introduced "Sites," and the two features are pointedly different. The source frames OpenAI as building a platform-as-a-service for durable, full-stack web applications, and Anthropic as building a stateless canvas for ephemeral, security-first visualization. The distinction matters more than the marketing race suggests, because the architectural constraints Anthropic accepts determine which enterprise use cases the feature can and cannot replace.

The basic mechanics of Artifacts are simple. According to Anthropic, an engineer working in Claude Code can ask the agent to spin up a webpage that pulls from the local repository, connected monitoring tools, and the conversational context already in the session. The page lives at a single URL the engineer can share with teammates, who must be authenticated members of the same organization to view it. As the agent continues to work in the terminal, the open page refreshes in-place: charts and text update at the same URL, and every revision lands in a version history that supports rollback and audit. The result is a shared, real-time view of what the agent is doing, which is the explicit value proposition the company is selling to engineering managers who have grown tired of status updates.

What the source does not sell is the cost of that simplicity. Anthropic's own documentation is unusually direct about the limits. Each artifact is a single, self-contained HTML page capped at a rendered size of 16 MiB. The published file is wrapped in a strict Content Security Policy that blocks all external network requests: no external scripts, fonts, or stylesheets; no fetch, XHR, or WebSocket calls. All CSS and JavaScript must be inlined, and images must be embedded as data URIs. Artifacts cannot store form input, call an API at view time, or serve multiple routes. The company summarizes the position plainly: "An artifact is a capture of work, not an application."

That line is the architectural thesis, and it sharpens when read against OpenAI's Codex Sites. According to the source's reading of the platforms' developer documentation, Sites is designed to generate durable, full-stack web applications that output as Cloudflare Worker-compatible ES modules. Agents on Sites can wire up D1 relational databases for structured data and R2 object storage for file uploads. A Site can support public sign-ins, integrate with external identity providers, and apply access controls tailored to specific workspace groups. The publishing flow is two-stage: a reviewable candidate linked to a Git commit, then a deploy to production. The product is positioned to replace functional internal SaaS tools, not to visualize the work of building them.

Anthropic's framing inverts that. Artifacts are not meant to be software at all, in the company's telling; they are meant to replace whiteboard diagrams, manual bug walkthroughs, and status reports with secure, self-updating visual tools that never leak live data outside the corporate boundary. The CSP that blocks all external network calls is the load-bearing piece of that promise, because it means an artifact cannot exfiltrate data to a third-party endpoint or pull in unvetted dependencies at view time. The 16 MiB cap, the inlined assets, the absence of multiple routes, and the lack of form persistence all serve the same goal: the published page cannot quietly turn into a real application with a real attack surface.

Both products are closed-source. Neither Claude Code Artifacts nor Codex Sites can be independently forked or self-hosted, and enterprise clients do not retain code-level ownership over the rendering engines or integration nodes. The source says both vendors have prioritized organizational security to make that vendor-managed approach palatable. Anthropic's controls are described as default-private artifacts viewable only by authenticated organization members, with org-level toggles, role-based scoping, explicit retention policies, and a centralized compliance API. OpenAI's controls are framed as gated by ChatGPT Business and Enterprise workspaces, with admin-enabled Sites access through RBAC, plus a Sites panel for managing runtime environment variables and secrets so keys do not have to live in source files. OpenAI's sharing model is slightly more granular (self and admins, all active users, or custom groups); Anthropic keeps the model narrower and the boundary tighter.

The internal perspective surfaced in the article comes from Boris Cherny, identified as the lead and creator of Claude Code, who described using artifacts for visual explanations, system diagrams, animation previews, and team-shared dashboards. The launch materials also walk through a worked example: an engineer prompts Claude Code to investigate user drop-offs after a release, the agent runs an SQL read, builds a drop-off funnel, identifies that "Pro accounts stall at the export sheet," proposes UI fixes, updates the live charts as code is refactored, and sends the manager a secure link that opens on mobile. That story is the cleanest demonstration of what Artifacts is for, and it also shows the limit: a single-page diagnostic tied to a specific session, not a durable app that retains state across users.

Anthropic's pattern clearly works for the visualization use case the company describes. The architectural restrictions that make the pattern work also set the ceiling, and the ceiling shows up in specific places. The CSP that prevents data exfiltration also prevents loading a corporate design system, embedding a third-party charting library, or calling an internal API to refresh a number the model cannot compute from session context. The 16 MiB cap rules out most realistic enterprise dashboards, and the absence of form input rules out interactive configuration. Teams that need any of those things will treat Artifacts as a starting point, not a destination, and will build the durable layer on top of it. The source frames Artifacts as a deliberate alternative to the Sites approach, and the architectural restrictions are the price of that positioning.