Connect AI Agents
Use Cremi's generation tools from Claude, Cursor, Codex and any MCP-compatible agent — or script them with the Cremi CLI.
Cremi exposes a hosted MCP (Model Context Protocol) server. Connect it to your favorite AI agent and it can generate with your account — credits, history and assets work exactly like in the app.
1.Get your access token
Every request is authenticated with a personal access token tied to your Cremi account.
- Open your profile menu and go to the AI Agents tab.
- Switch the toggle on to generate your token.
- Copy the token and keep it secret — anyone who has it can generate with your credits.
Regenerating the token or turning MCP off immediately invalidates the previous token.
2.Connect via MCP
Point any MCP client at the server URL below, sending your token in the mcp-access-token header.
https://cremi.ai/mcp-server/http
Run this in your terminal:
claude mcp add cremi --transport http https://cremi.ai/mcp-server/http \ --header "mcp-access-token: YOUR_ACCESS_TOKEN"
Replace YOUR_ACCESS_TOKEN with the token from step 1.
3.Available tools
Tools run against your account: credits are charged per generation at the same rates as the app, failed generations are refunded automatically, and results are saved to your Assets.
Generates an image from a text prompt, optionally guided by reference images for image-to-image and editing workflows.
| Model | Notes |
|---|---|
| imagen-4-fast · imagen-4-standard · imagen-4-ultra | Google Imagen 4 — fast, standard and ultra quality tiers |
| flux-pro · flux-pro-ultra | FLUX 1.1 Pro and Pro Ultra |
| openai · gpt-image-2 | OpenAI gpt-image-1, and GPT Image 2 — 1K / 2K / 4K resolution, up to 16 reference images |
| nano-banana · nano-banana-pro · nano-banana-2 | Nano Banana family — editing and reference images, 1K / 2K / 4K resolution |
Aspect ratios: 16:9 · 3:2 · 4:3 · 1:1 · 3:4 · 2:3 · 9:16
Resolutions: 1K, 2K, 4K (Nano Banana family and GPT Image 2 only)
Generates a video from a text prompt — optionally from a first and last frame, or from reference images, audio and video. Credits are charged on submission and refunded automatically if the render fails.
| Model | Notes |
|---|---|
| grok-imagine | Cheapest option — 720p, 1–15s |
| kling-2.1-standard · kling-2.5-turbo-pro | Kling 2.1 Standard and 2.5 Turbo Pro — 5s or 10s |
| veo-3.1-fast-silent · veo-3.1-fast · veo-3.1 | Google Veo 3.1 — silent, fast and full quality tiers; the non-silent tiers add a native audio track. 4 / 6 / 8s |
| seedance-2.0 · seedance-2.0-fast · seedance-2.0-mini | ByteDance Seedance 2.0 family — 2.0 at 480p / 720p / 1080p, Fast and Mini at 480p / 720p, 4–15s. Choose with resolution; the default is 720p. Takes up to 9 reference images, 3 audio and 3 video clips; references replace the first/last frame, and audio needs an image or video beside it |
| seedance-2.5 | ByteDance Seedance 2.5 — 480p or 720p, 4–15s. image_url and up to 9 reference images steer subject and look rather than setting the first frame. Also takes up to 3 audio and 3 video references |
| minimax-h3 | MiniMax H3 — 720p or 2K, 4–15s, native audio. Needs image_url or up to 5 reference images, used for subject and look rather than the first frame. Reference audio (native lip-sync) or video moves the render to MiniMax's own API, which is priced higher |
| minimax-h3-max | MiniMax H3 Max — the faster H3: 480p or 720p, 5–15s, native audio. Up to 2 reference images (image_url counts as one), 3 audio and 3 video references |
| wan-2.6-flash | Wan 2.6 Flash — image-to-video only, so image_url is required. 5 / 8 / 10 / 15s |
| wan-3 | Wan 3.0 — 480p / 720p / 1080p, 4–15s. Up to 10 reference images, 5 audio and 5 video references; images are references, not the first frame. With reference audio the clip keeps a soundtrack |
| pixverse-v6 · pixverse-c1 | PixVerse V6 and C1 (C1 is the cinematic tier) — 720p or 1080p, 4–15s, no audio track. Needs image_url or up to 7 reference images, which the video is built from; cite them in the prompt as @Image1, @Image2 … |
Aspect ratios: 16:9 · 9:16 · 1:1
Video takes minutes, so this tool submits the job and waits briefly. If it's still rendering you get a task id to poll — the render continues on our side either way.
reference_image_urls · reference_audio_urls · reference_video_urls · quote_only
reference_image_urls, reference_audio_urls and reference_video_urls take public URLs. Audio is .mp3 or .wav and video .mp4 or .mov, 2–15s per clip and at most 15s of each in total. Reference video — and reference audio on wan-3 — is billed for its own length as well as the output, so pass quote_only=true first to see the exact price without being charged. A model that cannot use an input rejects it before charging; it is never silently ignored.
Native lip-sync: give minimax-h3 the singer as a reference image and the vocal clip as reference_audio_urls — the same model Cremi Studio uses for singing scenes.
Turns a local image, audio or video file into a public URL in your Assets, so it can be passed to the frame and reference fields. upload_image does the same for images only. The CLI uploads local files for you.
Collect a video once it finishes rendering, and list the video models with their current credit cost, durations and resolution. Always read costs from list_video_models rather than assuming — pricing changes.
More tools — lipsync and motion control — are on the roadmap.
Prefer the terminal, or want to keep your coding agent's context lean? The Cremi CLI wraps the same tools. Cremi CLI guide →
Troubleshooting
- 401 Unauthorized / authentication failed
- Your token is missing, mistyped or was regenerated. Copy a fresh token from the AI Agents tab and update your client config. Use the hyphenated header name mcp-access-token.
- Insufficient credits
- Generations cost credits exactly like in the app. Top up or upgrade your plan, then retry — failed generations are refunded automatically.
- Generation times out
- Image generation can take a few minutes. Keep the default timeouts (the server sends progress updates while it works), and retry once if your network dropped.
- Tools don't show up in my agent
- Restart the client after editing its config, check the JSON/TOML syntax, and verify the server URL. In Claude Desktop, check Settings → Developer → MCP Logs.
Security notes
Your access token grants full generation access on your account and spends your credits. Treat it like a password: don't share it and don't commit it to repositories.
To revoke access at any time, turn the MCP toggle off or regenerate the token in Settings.