Krea Opens 60-Plus AI Image and Video Models to Coding Agents
Krea's hosted MCP server lets Claude, Cursor, Codex, and other agents generate images, video, and upscales from plain prompts.
- Krea launched a hosted MCP server at
api.krea.ai/mcpfor any agent client. - Works with Claude Code, Claude Desktop, Cursor, Codex, and OpenClaw via OAuth or API token.
- Exposes 60+ image and video models including Krea 2, Flux, Veo, Kling, Runway, Nano Banana Pro, Topaz.
- Tools cover model discovery, schema inspection, generation, uploads, polling, and job cancellation.
- Uses existing Krea compute credits, no separate agent billing to manage.
- 3D generation available via MCP but not yet in the web agent; Nodes workflows remain separate.
Krea opens its image and video stack to MCP agents
Krea has launched a hosted Krea MCP server that gives compatible AI agents access to its image, video, editing, workflow, and upscaling tools. Developers can generate media from Claude Code, Cursor, Codex, and other MCP clients without opening the Krea web app.
Model Context Protocol gives an AI client a standard way to discover and call external tools. The client model interprets a natural-language request, invokes Krea’s tools with structured arguments, and returns the results to the same session. Teams can add creative generation to existing agent workflows without maintaining a separate integration for every model provider.
One endpoint for 60-plus models
| Item | Details |
|---|---|
| Endpoint | https://api.krea.ai/mcp |
| Transport | Streamable HTTP |
| Authentication | Browser-based OAuth or an API token sent with Authorization: Bearer |
| Billing | Existing Krea compute credits |
| Job handling | Submission, status polling, result retrieval, and cancellation |
Krea’s router selects models according to the request and exposes live schemas that describe each model’s accepted inputs. The catalog includes Flux, Veo, Kling, Runway, Krea 2, Nano Banana Pro, Imagen, Sora, Wan, and Topaz. Topaz upscaling supports output resolutions as high as 22K.
Seven tools cover the job lifecycle
| Tool | Purpose |
|---|---|
list_models |
Return the image, video, and enhancement models currently available. |
get_model_schema |
Inspect a model’s current parameters and input requirements. |
generate |
Submit a generation, editing, or enhancement job. |
execute_node_app |
Run an existing Krea node application. |
get_job |
Check status and retrieve completed results. |
cancel_job |
Cancel an unwanted job. |
get_upload_url |
Create an upload destination for local or reference assets. |
A typical client lists the available models, fetches the relevant schema, uploads any reference files, submits the job, and polls until the output is ready. Images can return within seconds. Video and high-resolution jobs run asynchronously, allowing the client to continue other work and check their status later.
Connect Claude Code or Cursor
Claude Code with an API token
export KREA_API_TOKEN="your-token"
claude mcp add --transport http krea-ai \
https://api.krea.ai/mcp \
--header "Authorization: Bearer ${KREA_API_TOKEN}"The token comes from the Krea dashboard and grants access to billable generation. Store it in an environment variable or secret manager rather than a tracked configuration file.
Cursor with browser OAuth
{
"mcpServers": {
"krea-ai": {
"url": "https://api.krea.ai/mcp"
}
}
}Save the configuration as .cursor/mcp.json, restart Cursor, and complete the OAuth prompt when the client first connects. The same hosted endpoint works with Claude Desktop, Codex, OpenClaw, and other clients that support Streamable HTTP MCP servers.
Existing credits cover agent runs
Krea does not charge a separate agent subscription for MCP usage. Every generation draws from the same compute-credit balance used by the web application. Automated retries, large batches, video generation, and repeated upscales can consume credits quickly, so production clients should cap iteration counts and monitor long-running jobs.
Creative loops suited to MCP
- Product and campaign work: Generate product shots, campaign concepts, or storyboard frames during a coding or planning session, then revise the selected output.
- Video pipelines: Create a source image, animate it with Veo or Kling, and send the result through a Topaz upscale.
- Batch enhancement: Upload a folder of low-resolution assets and process each file through a controlled agent loop.
- Reference-driven series: Supply product or style references to maintain visual consistency across multiple outputs.
- Application prototyping: Generate interface art, placeholder media, and marketing assets from the same environment used to build the product.
Boundaries before automation
- 3D availability: MCP and Krea’s Slack agent expose 3D generation. The web agent currently omits it.
- Node authoring:
execute_node_appcan run an existing node application, but graph design remains a separate workflow. Teams that depend on custom ComfyUI-style pipelines still need their graph-authoring environment. - Variable latency: Video and large enhancement jobs take longer than typical image generations, so clients need polling and timeout logic.
- Credit exposure: An API token can trigger billable work. Applications should limit retries, batch sizes, and autonomous iteration.
- Client requirements: Hosts need Streamable HTTP support and a compatible OAuth or authorization-header flow.
A stable interface for changing models
The host model handles planning and tool selection; Krea executes the media work and manages its model catalog. Because clients can query available models and fetch live schemas, developers can build against one MCP interface as Krea adds or updates generation backends. That reduces provider-specific integration work and brings image and video production into the same agent session as code, research, and campaign planning.
Comments
