Artlist MCP

Official remote MCP for Artlist's AI Toolkit: image and video generation across 100+ models, invoked from inside a Claude conversation or agent run. Assets return inline and auto-save to a session library in the Artlist account.

Ground truth (verified 2026-07-10)

• Endpoint: https://mcp.artlist.io/mcp — remote streamable-HTTP MCP, OAuth-protected (unauthenticated → 401). Launched by Artlist 2026-07-06.
• There is no open-source Artlist MCP. GitHub code search for the endpoint returns nothing; no repo, npm package, or awesome-mcp listing implements one; Artlist publishes no official GitHub org. Community "artlist" repos are downloaders/scrapers, several ToS-violating. Never build or adopt a bespoke Artlist server — connect the official one. Artlist exposes no public generation REST API; the MCP is the sanctioned programmatic surface.
• Billing: paid Artlist accounts with AI credits. Every generation consumes credits; MCP usage is excluded from unlimited-generation plans. Treat every call as real spend.
• Not in the Claude connector directory — it is added as a custom connector. See references/connect.md.

How to use it

1. Discover the live catalog every session. The model roster (100+ and constantly changing — new generations land monthly) is exposed only through the server's live tool list. Fetch it fresh at the start of every session via ToolSearch; never hardcode mcp__… prefixes, and never reuse a tool list cached or carried over from a prior session — a stale list is a caching bug, so re-list. The live list is the sole source of truth and always overrides any model named in this file. Default to the newest generation the list currently exposes for a capability, dropping to a cheaper or faster tier only when the brief asks for it.
2. Brief, don't prompt-engineer. Describe the deliverable (subject, mood, format, market); the server maps intent to model parameters. Ask the MCP itself for a model recommendation when the brief is ambiguous — its answer reflects the current catalog, this file does not.
3. Iterate in-thread. brief → 2–3 directions → review inline → pick one → refine ("darker", "wider", "more movement") → localize variants. Context persists; never re-brief from scratch.
4. Match model to brief-part (live list decides). These names are illustrative as of 2026-07 and go stale fast — confirm each against the live list and take its current equivalent: stills/hero → Nano Banana Pro or Artlist Original 1.0; motion/product video → Seedance 2.0, Kling 3.0, or Veo 3.1; multilingual dialogue → Seedance 2.0 (native lip-sync). If a named model is absent, the catalog has replaced it with a newer generation — use that, or ask the MCP to recommend.
5. Retrieval. Every generation lands in a dedicated session in the Artlist account, searchable by description — cite the session when handing assets on, so nothing is re-generated (re-generation = duplicate spend).