Discourse AI - Web Artifacts

Documentation Site Management
1

What Are AI Artifacts?

AI Artifacts are powerful tools that allow users to create, embed, and interact with dynamic web-based components directly within a Discourse post. These components can contain custom HTML, CSS, and JavaScript, enabling a variety of use cases such as:

  • Embedding interactive quizzes or forms.
  • Visualizing data with rich graphics and animations.
  • Integrating lightweight web applications or tools.
  • A tool for learning about web frameworks, JavaScript libraries and more.

AI Artifacts seamlessly enhance Discourse posts by adding interactivity while ensuring a secure browsing experience.

Site Settings

Administrators can configure several settings:

1. Enabling the Feature

  • Setting: discourse_ai_enabled
    • Ensure this global setting is enabled for AI Artifacts to function.

2. Security Modes

  • Setting: ai_artifact_security
    • Options:
      • disabled: Disable the artifact system
      • lax: Artifacts automatically appear in posts without requiring user interaction.
      • strict: Users must explicitly activate artifacts in their browser by clicking a “View” or “Run” button. This setting is recommended for security-conscious environments. (default)

3. Artifact Creator Access

  • By default, the Artifact Creator persona is restricted to staff users only. This restriction ensures that only trusted individuals can create artifacts, minimizing the risk of improper or malicious usage.
  • If broader access is required, permissions need to be manually configured by a site administrator.

Using AI Artifacts in Posts

Artifacts are automatically created by the Artifact Creator persona on demand.

The above artifact was generated using GPT-4o and the anime.js library

Once created they are private to you and the Artifact Creator persona. Users who are able to share ai conversations (those who belong to the ai bot public sharing allowed groups) can also make an artifact public by sharing the conversation.

Once shared you can use the following HTML markup to render it in a post:

<div class="ai-artifact" data-ai-artifact-version="3" data-ai-artifact-id="71"></div>

(where the version and artifact id are your versioned artifact)

Security Considerations

Given that AI Artifacts can execute custom-written HTML, CSS, and JavaScript, Discourse has implemented robust safeguards:

1. Sandboxing in iFrames

  • Artifacts are rendered inside isolated iframe containers with sandbox attributes to restrict potentially unsafe operations like:
    • Cross-site scripting.
    • Access to external sites or APIs.

2. Content Security Policy (CSP)

3. Security Mode

  • Strict Mode: This is recommended for environments where artifacts are not fully trusted. Users will need to manually activate artifacts in their browser before they are rendered.

4. Access and Permissions

  • Artifacts are visible only to:
    • Their creator.
    • Users with the right permissions to view the associated post (e.g., private messages).
  • Public artifacts must be explicitly marked as such by sharing the AI conversation.

5. Length Limits

  • The size of HTML, CSS, and JavaScript in an artifact is capped at 64 KB each. This ensures components remain lightweight and do not burden users or systems.

Artifact storage

Web artifacts can optionally store per user data. To do so hint the artifact creator with “use user storage” or similar.

This system allows storing key,value pairs:

  • Private (only visible to admins and specific users)
  • Public (visible to all users including anonymous)

Key value pairs are secured against the post the artifact originated from, however if you share an artifact publicly all will be allowed to add keys.

To control storage you can use the hidden settings:

  • ai_artifact_kv_value_max_length (default items may only be 5000 letters or shorter)
  • ai_artifact_max_keys_per_user_per_artifact (default 100)

FAQ

Who Can Create AI Artifacts?

By default, only staff users (e.g., admins or moderators) can create artifacts through the Artifact Creator Persona. This persona simplifies the process of designing interactive web widgets using HTML, CSS, and JavaScript.

What Happens If I Click an Artifact?

  • In lax mode, artifacts appear automatically.
  • In strict mode, clicking the “Run” button activates the artifact and allows it to load in your browser.

Are AI Artifacts Safe?

Yes. AI Artifacts run in tightly controlled environments:

  • They are sandboxed and cannot interact with the Discourse application or user context directly, only via iframe messaging.
  • The strict mode gives you control over activation.
  • Artifacts are private by default you need to actively share to give global access.

Can I see the source code for Artifacts

Yes. When Discourse AI generates artifacts it will include the full markup, css and JavaScript.

What LLMs are supported?

Artifacts can be generated using any LLM you have configured, however various LLMs specialize and are more finely tuned for artifact creation.

We have seen good results with o3, Anthropic Claude Sonnet 3.7 - 4.0, GPT-4.1, Gemini Pro 2.5 and more. Generally more advanced models will do better.

There is a large aspect of trial and error with artifact creation, experimentation is key.

21 Likes
3

It’d be really neat to have slightly more granular control over artifact privacy so that they can be shared only with certain groups. Quite similar to how category permissions currently work :smiley:

4

I think that if you add the group to the bot pm it may work?

1 Like
5

I’m gonna test it out and report back here. I have a niche use case that requires artifacts to be accessible only to one group in one private category

1 Like
6

So I tagged the Web Artifact Creator bot in a new topic in a private category which the group has access so it can make an artifact only viewable to me and that particular group. However the artifact iframe window only shows that default " Oops! That page doesn’t exist or is private." that pops up when a user doesn’t have access to a particular page

So I made a data explorer query to check who could view that topic and the users in that group that should be able to see the artifact are indeed shown to be able to view it according to the database.

So perhaps this is a bug?

My SQL query 
-- [params]
-- int :artifact_id = 22

WITH artifact_info AS (
  SELECT 
    a.id,
    a.user_id as creator_id,
    a.post_id,
    p.topic_id,
    t.category_id,
    t.archetype,
    c.read_restricted,
    t.title as topic_title
  FROM ai_artifacts a
  LEFT JOIN posts p ON a.post_id = p.id
  LEFT JOIN topics t ON p.topic_id = t.id
  LEFT JOIN categories c ON t.category_id = c.id
  WHERE a.id = :artifact_id
),
users_with_access AS (
  -- Creator always has access
  SELECT 
    ai.creator_id as user_id,
    'Creator' as access_reason
  FROM artifact_info ai
  
  UNION
  
  -- Users with access to private messages
  SELECT 
    tau.user_id,
    'Private Message Access' as access_reason
  FROM artifact_info ai
  JOIN topic_allowed_users tau ON ai.topic_id = tau.topic_id
  WHERE ai.archetype = 'private_message'
  
  UNION
  
  -- Group members with access to private messages
  SELECT 
    gu.user_id,
    'Private Message Group Access' as access_reason
  FROM artifact_info ai
  JOIN topic_allowed_groups tag ON ai.topic_id = tag.topic_id
  JOIN group_users gu ON tag.group_id = gu.group_id
  WHERE ai.archetype = 'private_message'
  
  UNION
  
  -- Users with access to restricted categories
  SELECT 
    gu.user_id,
    'Category Group Access' as access_reason
  FROM artifact_info ai
  JOIN category_groups cg ON ai.category_id = cg.category_id
  JOIN group_users gu ON cg.group_id = gu.group_id
  WHERE ai.read_restricted = true
    AND ai.archetype != 'private_message'
    AND cg.permission_type IN (1, 2) -- full access or create/reply/see
  
  UNION
  
  -- All users if topic is public (not restricted and not private message)
  SELECT 
    u.id as user_id,
    'Public Access' as access_reason
  FROM artifact_info ai
  CROSS JOIN users u
  WHERE (ai.read_restricted = false OR ai.read_restricted IS NULL)
    AND (ai.archetype != 'private_message' OR ai.archetype IS NULL)
    AND u.active = true
)

SELECT 
  u.id as user_id,
  u.username,
  u.name,
  u.trust_level,
  uwa.access_reason,
  ai.topic_title
FROM users_with_access uwa
JOIN users u ON uwa.user_id = u.id
CROSS JOIN artifact_info ai
WHERE u.active = true
ORDER BY u.username