2dai-mcp-server

MCP server for 2DAI — lets Claude, Cursor, Cline, Windsurf and any other MCP host generate images and video, upload references, browse the public feed, and organise your cloud drive on your own 2DAI account.

• Runs locally over stdio via npx — nothing to install permanently.

• Authenticates with a 2DAI API key (create one at 2dai.io → Dashboard → Integrations → API keys).

• Finished images come back as inline previews so the model can see what it made and iterate.

New in 2.2 — Gen8 on the default channels

generate_image and generate_video run the Gen8 engines on their default channel; get_account names the engine behind each channel and what it takes (platform.engines, today Gen8.1):

• Images: up to 6 refCreationIds kept natively, legible text in the picture, long literal prompts honoured.

• Video: clips with sound, from a first frame plus up to 6 refCreationIds (characters, props and places kept consistent), an optional aspectRatio and an optional audio reference (upload_audio an MP3, then pass its id as audioCreationId, audioUse: "music" for a soundtrack or "voice" for a voice sample). Lengths are exact seconds — 1, 5, 8, 10, 12, 15, 17 or 20, opened by tier; video opens at Explorer.

• Presets and lengths depend on the tier: get_account lists what the account may submit and the best pick open to it (recommendation), so read it instead of probing.

next (imageModel / videoModel) is kept for the next engine. The next channel closed on 2026-11-11: it answers VIDEO_MODEL_UNAVAILABLE / IMAGE_MODEL_UNAVAILABLE, uncharged, until an engine opens it. Leave both parameters unset, and reconnect the server in your host after updating (a running session keeps the old tool descriptions).

Related MCP server: Krea MCP Server

New in 2.0 — reference tools on the Gen8 image engine

generate_with_refs and generate_in_artistic_style now render on the Gen8 image engine, which reads every reference natively: character-ref takes 1–6 shots of one character (face photos lock the exact same face, the prompt drives outfit, pose and scene), style-transfer 1–6 style sources, smart-edit the image to edit plus up to 5 support refs, the artistic style up to 6 subject images. The former face-ref tool was merged into character-ref and is refused with INVALID_TOOL. Same preset ids and tier gates; prices follow the Gen8 image ladder plus a surcharge per reference (the ticket's costUsd is the truth).

Quick start

Claude Code

claude mcp add 2dai --env TWODAI_API_KEY=2dai_sk_... -- npx -y 2dai-mcp-server

Claude Desktop / Cursor / Cline / Windsurf (JSON config)

{
  "mcpServers": {
    "2dai": {
      "command": "npx",
      "args": ["-y", "2dai-mcp-server"],
      "env": { "TWODAI_API_KEY": "2dai_sk_..." }
    }
  }
}

The key is read from the environment only — never pass it as a tool argument.

Hosted alternative — mcp.2dai.io:800

Every host that speaks remote MCP (Streamable HTTP) can point at our hosted server instead of running npx locally. No install, no Node runtime, no process lifecycle — just a URL and a Bearer key.

Endpoint — https://mcp.2dai.io:800/mcp

Auth — Authorization: Bearer 2dai_sk_... on every request (same key you'd put in TWODAI_API_KEY for the stdio flow).

Claude Desktop / Cursor / Cline / Windsurf (remote-MCP JSON shape — check your host's docs, some hosts still spell the type "remote"):

{
  "mcpServers": {
    "2dai": {
      "type": "streamable-http",
      "url": "https://mcp.2dai.io:800/mcp",
      "headers": { "Authorization": "Bearer 2dai_sk_..." }
    }
  }
}

Claude Code

claude mcp add --transport http 2dai https://mcp.2dai.io:800/mcp \
  --header "Authorization: Bearer 2dai_sk_..."

Both transports (stdio via npx and hosted via HTTP) expose the exact same tools. Pick whichever fits: npx when you want zero third-party dependency and full control of the process, hosted when you want zero install and a single URL.

No filesystem on the hosted server. The process runs in our container, so a savePath would write to our disk and an upload path would read from it — both are refused with a clear message (TWODAI_FILE_ACCESS=none is the hosted default). Instead, download_creation returns an inline preview (images) plus downloadUrl, the full-resolution file on the API: GET it with the same Authorization: Bearer header you send to the server (public creations need none). upload_image / upload_audio take base64. GET /health reports fileAccess: "none".

Tools

All wallet tools are read-only — no tool on this server can move money. The finance scope is opt-in and never part of a key's defaults; enable it in the dashboard's key settings when you want an agent to see wallet data.

Least privilege

Scopes are set per key in the dashboard. A coding agent that only generates needs read + generate; add manage for drive organisation, publish only if you want it posting to the public feed, and finance only for wallet reads. You can also set a per-key spend cap — the server reports actionable errors when a cap or scope blocks a call.

Configuration (env)

Notes for agents

• Presets. Image: max is the recommended everyday pick (best balance of detail and price); ultimate is the highest resolution for final masters; fast / normal are for drafts. Video: high at 5 seconds is the recommended shot; ultimate is the highest resolution with the longest wait. Every preset above normal (image) or fast (video) is tier-gated — get_account lists what the connected account may submit and the best pick open to it (recommendation), so never probe presets.

• Channels and engines. get_account names the engine behind each image and video channel and what it takes (platform.engines: a summary sentence plus, from server 2.9.0, the reference cap maxRefCreationIds, sound, aspect ratio and frame rate); next (imageModel / videoModel) is the channel kept for the next engine — leave it unset unless get_account shows a distinct engine there. A closed channel is refused, uncharged.

• Lengths. A clip runs at the length asked, in exact seconds: get_account lists the lengths the account may submit (videoDurations, tier lock resolved); any other length is refused with INVALID_DURATION, uncharged.

• Images keep up to 6 refCreationIds natively, renders legible text and honours long literal prompts — describe the scene plainly rather than in keyword lists. Same quality ids, priced on render time, and every reference adds a surcharge. get_account reports the presets open to the account (qualities, canUseImageNext for the next channel); a submit below a gate is refused with the tier that opens it, uncharged.

• Video renders clips with sound at 24 fps from a first frame plus up to 6 refCreationIds (each adds a surcharge), with an optional aspectRatio (auto keeps the first frame's ratio) and an optional audio reference — an MP3 from upload_audio as audioCreationId, audioUse: "music" (soundtrack / sound design) or "voice" (a voice sample the speaking character follows); it counts as one more reference. Lengths and presets depend on the account: get_account lists what it may submit. Long clips render for several minutes: expect a queueId and collect it with check_generation.

• Shots. With style: "raw", write the clip the way TIXI does: the first frame is the opening shot (describe it as it is at 0 s, then the changes in order), about one shot per 1.7–2 s (5 s = 2–3, 8 s = 4, 10 s = 5, 20 s = 12; under 3 s one continuous shot), one camera idea and about one action per shot, each subject anchored by its face, hair, garments and colours, and each sound written next to the action that makes it.

• Video styles and TIXI. Every video style except raw (and auto) runs TIXI on the prompt: the style's guidance only exists through that rewrite, so it cannot be switched off. style: "raw" sends the prompt verbatim — pass enhance: true to have TIXI write the clip's brief (shots, pictures and, on an engine with sound, the sound) from a short idea; a prompt starting with tixi adds a thinking pass (the word is dropped from what TIXI reads). No extra credit either way.

• Price — video is priced on render time, which grows faster than the length (8 s ≈ 1.7–2.2×, 10 s ≈ 2.2–3.4×, 20 s ≈ 6–12× the 5 s price, higher presets at the top of each range), +5 % per reference after the first frame, the audio reference included. auto draws the preset within the account's tier, so pin it for a predictable cost; every submit returns the resolved quality next to costUsd.

• Batches. check_generation takes queueIds (up to 25) to poll a whole batch of shots in one call.

• Prompts are never cut silently. Listing rows shorten prompt / description with an ellipsis and a promptTruncated flag; get_creation returns the full text, so a prompt can be re-read and replayed.

• References are checked at submit. An inputCreationId / refCreationIds entry that is unknown, in the trash, deleted, or not visible to the account is refused with CREATION_NOT_FOUND and the ids, before any charge (restore it with organise_creation, or drop it).

• Prompts are capped per account tier (1,500 characters as a Guest, 8,000 at Supporter, 20,000 at Founder): get_account reports the exact cap as promptMaxChars and the tier that raises it as promptMaxCharsNext; a longer prompt is refused with PROMPT_TOO_LONG before any charge, a longer negativePrompt is truncated to the cap. TIXI enhancement is off by default on generate_image (enhance: true rewrites a short brief into a full prompt); character-ref, wallpaper resize and the artistic-style tool always run it.

• Costs are charged at submit against the account's USD credit and refunded on failure; get_account shows the headroom, get_stats the 30 / 90-day burn.

• Content rating is not an error. Every output carries nsfwLabel / nsfwRate. From Near-nude (0.8) the reply adds contentRestricted: true: the generation succeeded and was charged, the creation is kept but masked in the owner's drive, cannot be published, and its caption is withheld (get_creation returns it). Hand over the viewUrl for review instead of retrying. Only NSFW_MAX_EXCEEDED discards an output, uncharged.

• Skill. The package ships skills/2dai/SKILL.md — a platform guide for agents (quality ladder, costs, the film recipe, brand-consistency recipe, limits). Point your agent runtime at it, or read it once per session.

The 2DAI stack

2DAI runs its own models, and unlike others, we don't resell model access. All our models originate from the open-source community or our own R&D — we then disassemble, modify, fine-tune and optimize them to align with our legacy and 2DAI❤️ART lines. They also run on our own private cloud network.

get_account names the engine behind each channel (platform.engines, modelLabels); the next channels are kept for the next engine.

Every tool response carries the model version and stack summary (via get_account.platform), so an agent can hand these facts back to a user who asks "what powers you?" without extra plumbing.

License

MIT

Community & links

• Website — www.2dai.io

• X — @2DAICommunity

• Telegram — t.me/Token2dAI

• $2DAI on Base — chart on Dexscreener · stake for tiers & daily credit airdrops

• SDK — 2dai-cloud-sdk for scripts, servers and apps

• GitHub — 2DAICommunity