| name | artlist-mcp |
| description | Use when a task needs AI image or video generation — campaign visuals, hero shots, product or social video, thumbnails, ad creative, localized variants — through the official Artlist MCP (mcp.artlist.io, 100+ models incl. Nano Banana Pro, Seedance 2.0, Kling 3.0, Veo 3.1, Sora 2). The model catalog is discovered live from the server every session, never a stored or cached list. Use also before any autonomous run that could touch media generation, because it carries the credit-spend gate. Covers connection per surface, live model selection, the iterate-in-conversation loop, and failure handling. |
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
- 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.
- 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.
- 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.
- 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.
- 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).
Governance — charter-bound
Generation consumes paid credits, so it sits under the spend rules of self-improvement-charter:
- Autonomous / unattended runs: generation permitted only when the run declares a pre-approved credit budget at start. On exhaustion: stop and report. An agent never raises its own cap; a denied spend gate is terminal.
- No unattended bulk loops. Batch generation (campaign matrices, per-market fan-outs) requires explicit human sign-off on batch size before the first call.
- Interactive sessions: confirm with the human before any burst larger than ~3 generations.
- Auth is per-human, per-surface. OAuth tokens are never exported, shared between fleet nodes, or committed. A 401 means re-authenticate through the official flow — nothing else.
- Licensing check before shipping. Confirm the account tier covers the intended commercial use before an asset leaves the boardroom or lands in client work.
Nexus wiring
Registered in agentskills.json; consumers (CCW-CRM boardroom, fleet nodes, Synthex) receive it through the live-fetched Nexus Prompt — never fork this file into a consuming repo, same rule as the Nexus Prompt. Boardroom runs count as autonomous runs and pass through the spend gate above. Pairs with other connectors in one conversation (Drive brand guides, Linear briefs): pull context first, then generate.
Failure modes
| Symptom | Meaning | Action |
|---|
| 401 from endpoint | OAuth expired or never completed | Re-auth via the surface's official flow; do not retry blind |
| Credit/quota error | Budget exhausted | Stop, report spend, await human — never retry-loop |
| Tool list looks stale / missing new models | Client reused a cached/prior-session list | Force a fresh tools/list; refresh or reconnect the connector — never select from a cached catalog |
| Named model absent | Catalog rotated to a newer generation | Re-list; the newer generation has replaced it — use that or ask the MCP to recommend |
| Connector missing on a fleet node | Node never authorized | Follow references/connect.md; auth is per-node, no fleet-wide token |
References
references/connect.md — connection steps for claude.ai/Desktop, Claude Code (fleet nodes), and the API, plus post-connect verification. Read when installing on a new node.
