Skip to main content

Artifacts

Artifacts are deliverables the agent produces during a conversation — files, websites, images, code, data, and more. Unlike plain-text responses, artifacts are versioned, previewable, and downloadable.

How artifacts are created

You do not create artifacts directly. The agent creates them by calling one of two tools:

  • save_artifact — registers a file or set of files the agent wrote (a website, document, code project, data file, or skill bundle) as an artifact.
  • generate_image — produces an image artifact directly from a prompt.

This happens naturally when you ask the agent to build something concrete:

  • "Build me a landing page"
  • "Generate a bar chart from this CSV"
  • "Write a Python script that processes these logs"
  • "Create a project proposal document"
  • "Draw an illustration of a mountain at sunset"

For save_artifact, the agent writes the files (using code execution in the sandbox), then registers them as an artifact so they appear in your conversation.

Artifact types

TypeWhat it isPreview behavior
WebsiteHTML/CSS/JS sites or appsLive rendered preview in an iframe
DocumentMarkdown, PDF, DOCX, or other text documentsRendered document preview (Markdown rendered, PDF embedded, Office documents via viewer)
ImagePNG, SVG, JPEG, and similarInline image preview
CodeSource code files or projectsSyntax-highlighted code viewer
DataCSV, JSON, Excel spreadsheetsTabular data preview
SkillA skill bundle with a SKILL.md entry pointSkill preview with option to publish to your workspace
FileAnything that does not fit the above categoriesDownload link with basic metadata

Viewing artifacts

When the agent creates an artifact, an artifact card appears inline in the chat message. The card shows:

  • The artifact name and type icon.
  • A brief description (if the agent provided one).
  • A version badge (for artifacts with multiple versions).
  • Preview (eye icon) and Download buttons.

Clicking the card (or the preview button) opens the artifact panel on the right side of the screen. The panel renders a live preview based on the artifact type — websites run in a sandboxed iframe, images display at full resolution, code gets syntax highlighting, and so on.

📸 Screenshot placeholder

Capture: An artifact card inline in the chat with the artifact panel open on the right, showing a live website preview and the version badge expanded into the version dropdown. Asset: /img/conversations/artifact-panel.png

At the top of the chat area, a collapsible Artifacts bar shows all artifacts created in the current conversation. Click to expand the gallery and quickly jump to any artifact without scrolling through the message history.

Versioning

The agent can update an artifact across multiple turns. When you say "make the header blue" or "add a footer to the page," the agent creates a new version of the existing artifact rather than a separate artifact.

Versioned artifacts show a version badge (e.g., "v3") on their card. In the artifact panel, click the version badge to open a dropdown listing all versions with timestamps. Select any version to preview it or download that specific version.

Downloading artifacts

Every artifact can be downloaded to your local machine:

  • Click the download icon on the artifact card in the chat.
  • Or click download in the artifact panel header.

For multi-file artifacts (like a website with HTML, CSS, and JS files), the download packages everything into a single archive.

Skill artifacts

When the agent creates a skill (via the skill-creator workflow), the artifact type is skill. Skill artifacts have a special action: Add to workspace. Clicking this publishes the skill to your workspace so it becomes available in future conversations (the agent can load it via load_skill). The version comes from the version field in the skill's SKILL.md frontmatter; to update a published skill, bump that version and add it again.

Tips

  • Be specific about what you want. "Build a dashboard" is vague. "Build a dashboard with three cards showing daily active users, revenue, and error rate using Chart.js" gives the agent enough detail to produce a useful artifact on the first try.
  • Iterate by describing changes. After the agent creates an artifact, ask for modifications in follow-up messages. The agent updates the same artifact with a new version, keeping your history clean.
  • Use the preview panel. Do not rely solely on the agent's text description. Open the preview to see exactly what was produced, especially for websites and images.
  • Download important artifacts. Artifacts live within the conversation. If you need the output long-term, download it to your local machine or a shared drive.