Open Artifacts

English | Русский | Landing Page | Changelog | История изменений

Self-hosted artifact hosting for AI agents — an open alternative to Claude Artifacts. Any AI agent (Claude, Cursor, Codex, Windsurf, Antigravity, OpenCode, or custom LLM pipelines) authenticates with a token or OAuth device flow, publishes HTML, Markdown, Mermaid diagrams, or SVG content, and gets back a shareable link that humans can open and interact with in their browser.

Runs as a lightweight, single-stack Docker Compose deployment.

🌐 Landing Page: https://emaxe.github.io/open-artifacts

Topics & Tags

ai-agents • artifacts • claude-artifacts • claude • mcp • mcp-server • model-context-protocol • self-hosted • docker • hono • react • typescript • cli • developer-tools • open-artifacts • llm • s3 • constructor • yaml • block-builder

Key Features

• 🛡️ Zero-Trust Security Sandbox: Untrusted agent-generated artifacts render inside strictly isolated <iframe> containers with sandboxed origins (allow-scripts, no credential or cookie leaks, strict Content Security Policy).

• 🧩 Constructor Mode (Declarative Artifact Builder): Fast, token-efficient assembly of HTML deliverables. Agents describe pages as clean YAML/JSON specifications or reference existing Markdown reports using 23 modular blocks. Compiles in milliseconds, cuts token generation costs by ~80%, and eliminates CSS/HTML hallucinations.

• 📦 Multi-Format Support: Interactive HTML applications, GitHub-flavored Markdown, responsive Mermaid diagrams, and raw vector SVG graphics.

• ⚡ Built-in MCP Server: Ready-to-use Model Context Protocol endpoint at /mcp (Streamable HTTP) for immediate integration with Cursor, Claude Desktop, and Claude Code.

• 🤖 Standard AI Agent Skill: First-class support for skills.sh (npx skills add emaxe/open-artifacts) with automatic CLI detection and device-flow authorization.

• 🎨 Built-in Design Systems: Four ready-made design templates (dashboards, documents, promo/decks, diagrams) ship with the skill so an agent picks the right one before publishing, keeping artifacts visually consistent.

• 🔑 Flexible Authentication: Interactive OAuth Device Flow (oa login) and non-interactive organization API tokens (OA_TOKEN).

• 🖥️ Bilingual Sign-In Screen: / greets a signed-out visitor with a two-panel sign-in/sign-up screen (English/Russian, toggle in the header) — no marketing copy to scroll past on your own instance. Signed-in visitors are redirected straight to their own workspace. The product's marketing site lives separately at emaxe.github.io/open-artifacts (source: apps/landing/), published to GitHub Pages and carrying no auth of its own.

• 👥 Multi-Tenancy & Teams: Organizations, user management, and fine-grained roles (superadmin, admin, member) with an intuitive Team Switcher.

• 🔗 Secure Sharing: Team-only links, password-protected links, public links (optionally labeled for your own reference), and automatic link expiration (1 hour, 1 day, 7 days, 30 days) — teams and instance admins control the default link mode and can disable public links entirely. Every share page shows a branded viewer panel: one compact row with the title, type, version and date. Anyone who can manage the artifact also gets a version picker, a collapsible details block (size, views, expiry, description) and a visibility control — pick Team only, Password protected or Public and it applies instantly, in place, on the same link: no page reload, no revoking it and handing out a new URL.

• 🖼️ File Attachments: Images and other files can be uploaded to an S3-compatible bucket (MinIO ships in docker-compose.yml by default) and attached to an artifact — proxied through the app, never a public bucket. An agent can check its remaining quota (oa quota / the get_storage_quota MCP tool) before deciding whether to upload. Instance-wide and per-team storage quotas (team-wide and per-artifact, unlimited by default) keep usage in check; files are deleted automatically with their artifact.

• 🚀 One-Command Deployment: Instant production setup with Docker Compose or the interactive ./run.sh runner.

Related MCP server: artifacty

Quick Start

Option 1: Interactive Runner (Recommended)

./run.sh

An interactive menu to orchestrate Docker production containers, local development, builds, tests, and database migrations without memorizing individual commands.

Option 2: Docker Compose

cp .env.example .env        # Configure SESSION_SECRET, SUPERADMIN_EMAIL/PASSWORD
docker compose up -d

Open http://localhost:3000, log in using the superadmin credentials configured in .env, and you're ready! Database migrations run automatically on container startup.

Project Layout

apps/api/               Hono REST API + MCP server + Postgres (Drizzle) + Static web server
apps/web/               React SPA (Vite) — sign-in screen + Admin & User interface
apps/landing/           Static marketing site (Vite, no backend) — published to GitHub Pages
packages/shared/       Zod schemas + shared validation rules (access control, CSP, TTL parsing)
packages/cli/          `oa` CLI for AI agents (published on npm as @emaxe/oa)
skills/open-artifacts/ Agent skill (SKILL.md) + design templates (references/design/) compatible with skills.sh
docker/                Multi-stage Dockerfile for containerized deployment
run.sh                 Interactive terminal management script

Local Development

Requirements: Node.js 20+, pnpm, and Docker (for local PostgreSQL).

# 1. Install dependencies
pnpm install

# 2. Start PostgreSQL dev database on localhost:5433
docker compose -f docker-compose.dev.yml up -d

# 3. Synchronize database schema
cd apps/api && DATABASE_URL=postgres://postgres:postgres@localhost:5433/open_artifacts_dev \
  pnpm exec drizzle-kit push

# 4. Start development servers
pnpm dev:api      # API & MCP server at http://localhost:3000
pnpm dev:web      # React UI at http://localhost:5173 (proxies /api to :3000)
pnpm dev:landing  # Marketing site at http://localhost:5174 (apps/landing, no backend needed)

Develop and test the CLI locally:

cd packages/cli && pnpm dev -- login --server http://localhost:3000

Tests

pnpm -r run test                                             # Run all unit tests
cd apps/api && pnpm run test:integration                     # Integration tests (requires dev DB)
cd apps/api && pnpm exec playwright install chromium          # Install Playwright browser once
cd apps/api && pnpm run test:e2e                              # End-to-end security sandbox tests

Security Verification: The e2e test suite (apps/api/e2e/sandbox-security.spec.ts) verifies sandbox isolation by injecting malicious payloads and confirming that artifacts cannot access host cookies, localStorage, or backend APIs.

AI Agent Skill (for Claude, Cursor, Windsurf, etc.)

Open Artifacts ships with an agent skill compliant with the skills.sh standard. It instructs AI assistants how to automatically install @emaxe/oa, authenticate, publish generated artifacts, and present clean preview URLs in chat. It also ships four ready-made design systems — dashboards, reports, promo/decks, and diagrams — plus a shared foundation, which the agent picks between before writing an html artifact, so published artifacts look consistent regardless of which agent or session produced them.

1. Install via skills.sh

# In the current project workspace (recommended):
npx skills add emaxe/open-artifacts

# Or install globally:
npx skills add emaxe/open-artifacts -g

2. Agent Authentication

Agents can authenticate using either of two methods:

Method A: Personal access (recommended)

1. Run in terminal or agent session:

   oa login --server http://localhost:3000

2. The CLI outputs a verification code (e.g. ABCD-1234) and activation URL.

3. Open http://localhost:3000/activate?code=ABCD-1234 in the browser and approve — no organization to pick here.

4. This issues a personal key, valid across every team the approver belongs to (with their real role in each), saved to ~/.config/open-artifacts/credentials.json (600 permissions). Pick a team for a given project with:

   oa orgs           # list your teams
   oa use <team>     # set the default for this project (writes .oa.json, no secrets, safe to commit)

Method B: Agent key locked to one team (for CI)

Run oa login --agent (organization selection required at approval time), or create an API key in the web UI under Team > Agents (/t/:orgId/agents) and pass it to the agent's environment:

export OA_SERVER="http://localhost:3000"
export OA_TOKEN="oa_live_xxxxxxxxxxxxxxxxxxxxxxxx"

Verify connection:

oa whoami

3. CLI Command Cheat Sheet

Supported file formats: html, markdown, mermaid, svg. Append --json to any command for structured JSON output.

🧩 Constructor Mode: Declarative Artifact Builder

Writing raw HTML/CSS from an AI agent has significant drawbacks:

• Massive token consumption: 2,000–5,000+ output tokens per artifact, consuming context windows and driving up costs.

• High latency: Streaming thousands of repetitive HTML tags takes 30–60+ seconds.

• Fragile styling & layout hallucinations: LLMs frequently introduce syntax errors, broken CSS, missing responsive styles, broken dark mode, or blocked CDN imports that fail the strict sandbox CSP.

Constructor Mode solves this completely. Rather than generating raw HTML, the agent emits a concise, structured YAML or JSON specification. A lightweight builder (skills/open-artifacts/constructor/build.mjs) then compiles it in milliseconds into a standalone, accessible, dark-mode-ready HTML artifact.

Key Benefits

How It Works

┌───────────────────────────┐      ┌───────────────────────────────┐      ┌─────────────────────────┐
│  AI Agent                 │      │  Constructor (build.mjs)      │      │  Open Artifacts Server  │
│                           │      │                               │      │                         │
│  Emits YAML/JSON spec     │ ───► │  • Validates blocks           │ ───► │  Self-contained HTML    │
│  or points to .md report  │      │  • Inlines CSS themes         │      │  Sandboxed iframe       │
│  (~80% fewer tokens)      │      │  • Dedupes Chart/Mermaid CDNs │      │  Shareable URL link     │
└───────────────────────────┘      └───────────────────────────────┘      └─────────────────────────┘

1. Compile & Publish in One Command

# Build and publish with an auto-generated share link:
node skills/open-artifacts/constructor/build.mjs report.yaml --share

2. Declarative Specification Example (report.yaml)

title: "Quarterly Analytics"
theme: data # default | data | document | promo | diagram
lang: en

blocks:
  - type: hero
    title: "Q3 Performance Dashboard"
    subtitle: "Enterprise telemetry & ARR"
    badge: "Verified"

  - type: kpi-row
    items:
      - label: "Total Revenue"
        value: "$4.25M"
        delta: "+18.4%"
        trend: up
        icon: "💰"
      - label: "Active Clients"
        value: "1,240"
        delta: "+5.1%"
        trend: up
        icon: "👥"
      - label: "Monthly Churn"
        value: "1.18%"
        delta: "-0.2%"
        trend: down
        icon: "📉"

  - type: two-columns
    ratio: "1:1"
    left:
      - type: chart-bar
        title: "Monthly ARR ($k)"
        labels: ["Jul", "Aug", "Sep"]
        datasets:
          - label: "Actual"
            data: [740, 780, 840]
            color: c1
    right:
      - type: chart-pie
        title: "Revenue by Tier"
        donut: true
        labels: ["Enterprise", "Mid-Market", "SMB"]
        data: [55, 30, 15]

  - type: tabs
    items:
      - label: "Accounts"
        blocks:
          - type: table
            caption: "Top Accounts by ARR"
            sortable: true
            columns: ["Client", "Tier", "ARR", "Status"]
            rows:
              - ["Acme Global", "Enterprise", "$420k", "Active"]
              - ["Globex Tech", "Enterprise", "$380k", "Active"]
      - label: "Architecture"
        blocks:
          - type: mermaid-diagram
            caption: "Processing Pipeline"
            definition: |
              graph LR
                A[Agent] --> B[Constructor]
                B --> C[Open Artifacts]

  - type: image
    src: "./assets/diagram.png" # local images auto-inline to base64
    alt: "System Architecture"
    caption: "Click to zoom"
    width: "400px"
    zoomable: true # Fullscreen lightbox on click

  - type: raw # Escape hatch: arbitrary HTML/CSS/JS with access to theme tokens
    html: |
      <div style="background: var(--surface-2); padding: 12px; border-radius: var(--radius);">
        Custom component styled with active theme tokens
      </div>

3 Ways to Publish Markdown Reports

Already have a finished Markdown document (research note, RFC, post-mortem, or audit)? You don't need to rebuild it as blocks:

1. Option A — Passthrough Mode (Lowest token cost): Point directly to the markdown file. The builder parses it, applies the document theme, and outputs a clean standalone artifact. Frontmatter (---...---) is stripped automatically:

   title: "Q3 Research Note"
   theme: document
   source: /path/to/report.md

2. Option B — markdown-file block (Framed report): Combine an existing .md file with top-level KPI metrics or a hero banner:

   title: "Executive Summary"
   theme: document
   blocks:
     - type: hero
       title: "Q3 Executive Summary"
     - type: kpi-row
       items: [...]
     - type: divider
       label: "Full Report"
     - type: markdown-file
       path: /path/to/report.md

3. Option C — markdown block (Inline): Embed Markdown prose directly inside the YAML specification:

   blocks:
     - type: markdown
       content: |
         ## Executive Summary
         Revenue grew **18% QoQ** driven by expansion in the enterprise segment.

23 Built-in Modular Blocks

Independent CSS Themes

Themes decouple aesthetics from content. Switching a theme takes one line in YAML:

• default: Clean neutral theme suitable for general documents and mixed deliverables.

• data: Optimized for dashboards, dense metrics, sortable tables, and hover-highlighting.

• document: Optimized for long-form reading (720px readable width, 1.8 line-height, elegant typography).

• promo: High-impact landing pages and announcements (gradient hero, elevated cards).

• diagram: Centered, responsive layout for architecture and flow diagrams.

Model Context Protocol (MCP) Server

Every Open Artifacts instance serves a built-in MCP server at <APP_ORIGIN>/mcp via Streamable HTTP. Connect any MCP client by supplying an API key (personal or agent) in the Authorization header. A personal key spans every team it belongs to — ?orgId= in the URL sets the default team for this connection (omit it if the key only has one team):

{
  "mcpServers": {
    "open-artifacts": {
      "url": "http://localhost:3000/mcp?orgId=<team-id>",
      "headers": {
        "Authorization": "Bearer oa_live_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Available MCP Tools

• whoami: Check token identity, scopes, and team context.

• list_orgs: List the teams this identity can act in.

• list_artifacts: Search and retrieve artifacts in a team.

• get_artifact: Fetch artifact metadata and raw content.

• create_artifact: Create a new artifact (html, markdown, mermaid, svg).

• update_artifact: Publish a new version for an existing artifact.

• delete_artifact: Soft-delete an artifact.

• create_share: Create a team-only, password-protected, or public expiring share link. Omitting mode uses the team's configured default.

• list_shares: View all active shares for an artifact.

• revoke_share: Instantly deactivate a share link.

list_artifacts and create_artifact take an optional orgId argument, overriding the connection's default. With neither set and more than one team to choose from, they return a structured org_required error listing the candidates instead of guessing.

Agent Instructions Page

An interactive cheatsheet with copyable commands, live instance URLs, and direct links to /activate is integrated into the web UI at /help/agents, available to every signed-in user — connecting an agent is something any team member does for their own team, not a superadmin-only task.

Teams, invites, and admin

Every user gets an auto-provisioned personal main workspace on registration (including when registering through an invite link — they land in both their own workspace and the inviting team). Teams are created explicitly from Команды (/teams).

Inviting someone to a team (/t/:orgId/settings) works for both new and existing accounts:

• New email — you get a one-time link (/invite/:token) to send them yourself; there is no email delivery built in.

• Existing account — they see the invite under Приглашения (/invites) in their own session and accept or decline it explicitly; nothing is added silently.

Superadmins get a dedicated admin area at /admin (Overview, Users, Teams, Audit, Instance settings), separate from the agent-connection instructions above.

Configuration

Configure the application through environment variables (see .env.example):

Most operational settings — registration mode, default key TTL, invite link lifetime, the CDN allowlist, the maximum artifact size, the maximum/default artifact lifetime, and storage quotas — can also be modified at runtime by superadmins in Настройки инстанса (/admin/settings); they don't need an environment variable or a restart.

File storage and quotas

Uploaded files (images, attachments) live in an S3-compatible bucket, proxied through the app at /af/<token> — the bucket itself stays private, no presigned URLs are ever handed out. A file always belongs to exactly one artifact and is deleted with it (soft-delete, expiry, or an org being deleted all queue its files for removal); there is no way for an orphaned file to exist.

Storage quotas are unlimited by default, at both the team and per-artifact level. A superadmin sets instance-wide ceilings in /admin/settings; a team can set its own stricter override (never looser) — the per-artifact one in its own team settings, the team-wide one only by a superadmin. Before uploading, an agent can check GET /api/v1/quota (or oa quota / the get_storage_quota MCP tool) to see remaining room and decide whether the upload is worth attempting.

Artifact lifetime (TTL)

By default artifacts live forever. A superadmin can set an instance-wide maximum artifact lifetime (in minutes, in /admin/settings) — it doubles as the default for newly created artifacts. Team owners/admins may set their own stricter limit for their team, but never looser than the instance maximum. Callers (CLI --lifetime, the web UI, or the MCP/API lifetime field) may pick anything up to the effective limit at creation time, and change it later.

Once an artifact's lifetime expires its content is hard-deleted — the version history is gone for good, only a tombstone row remains for audit/analytics. Lowering the maximum re-clamps every existing artifact's expiry from its own creation date (min(current deadline, created_at + new max)); raising it never extends anything already created.

Link modes and share policy

A share link is team (only logged-in members of the artifact's team can open it), password (anyone with the URL and the password), or public (anyone with the URL). Creating one without naming a mode — the plain --share flag, an omitted mode field over the API, or MCP's create_share with no mode argument — uses the team's configured default, not always public.

Team owners/admins set two things per team in Настройки команды (/t/:orgId/settings, including for their personal main workspace): the default link mode (team or public) and whether public links are allowed at all. A superadmin sets the same two knobs instance-wide in /admin/settings; a team can only be equal-or-stricter than the instance, never looser — it can never re-enable public links the instance disabled, nor default to public while public links are forbidden. Disabling public links only blocks new ones; existing public shares keep working until revoked, one by one from the artifact page or in bulk from the team settings page.

An existing link's own access mode can also be changed in place — without revoking it and handing out a new URL — right from its own viewer page (/s/:token), by anyone who can manage the artifact (PATCH /api/v1/shares/:id). Picking a mode in the panel's visibility menu saves it immediately and updates the panel without reloading the page; Password protected first asks for the new password.

License

MIT © emaxe