Krea Agents Replaces Complex Node Graphs With a Single Creative Prompt

Krea has opened the beta for Krea Agents, a layer on top of its creative suite that lets AI agents drive image and video production end to end. Describe what you want, and the agent reads the canvas, plans a multi-stage pipeline, picks the models, and runs the job from a single prompt.

Goodbye node graphs

Professional AI creative work has largely been trapped inside ComfyUI-style node graphs that only technically fluent users can operate. Krea Agents targets that friction directly. If Krea 2 pulled model quality in-house, Agents pulls the full production pipeline in-house, replacing hand-wired graphs with natural language.

The system builds on Krea's earlier Node Agent, which already turned prompts into runnable graphs. When you change something, only downstream nodes rerun. Swap a prompt at the top of a ten-node chain and the first nine cached results stay put while the affected nodes reprocess. That caching hints at how the new platform keeps iteration cheap.

Under the hood

Krea splits the stack in a way most creative agents do not. Planning and reasoning come from frontier LLMs, including Anthropic's models, while generation runs on Krea's own models plus whatever else is on the platform. You choose which LLM the agent uses, dial how much effort it should spend, and complex jobs get delegated to specialist sub-agents.

The orchestration surface is broad:

• Access to 60+ models including Flux, Veo 3.1, Kling, Runway, Luma, and in-house Krea 2.
• A right-side toolbar with media visualizers and file access, built to give the agent visual feedback.
• A file system for persistent context, which is how the agent learns your preferences and past work.
• Skills you can build from a set of reference images and invoke by tagging them in the prompt box.

Memory and taste as first-class objects

The most interesting departure from generic coding agents is the context manager. The file system stores what the agent has learned about your style, your brand, and the projects you keep returning to. Seed it by importing a moodboard from Pinterest, uploading references, or letting the agent build a skill from a batch of stylistically consistent images.

Skills work as reusable style capsules. Once trained, they appear as taggable entities in the prompt box, so instead of re-uploading references every session you drop a tag and the agent carries that visual direction through whatever pipeline it plans.

Plugs into tools you already use

Krea is betting on interop over lock-in, committing to MCP and API access so Agents slots into existing Claude Code or Cursor pipelines rather than replacing them. The Krea MCP server exposes the platform to external agents via OAuth.

Through MCP, agents can discover available image and video models, inspect live input schemas, submit generation jobs, upload local assets, poll results, and cancel mistakes. Inside Krea Agents itself, connectors reach Slack, Figma, and Google Drive, so the agent can pull project context, post outputs, or run an ad campaign across social channels.

Access and pricing

The beta is invite-only, distributed through the waitlist and social engagement on the launch announcement. To seed usage, Krea is offering free LLM calls and extra generations inside Agents for the first week. Reach it from the top-left corner of the main Krea app or directly at krea.ai/agent.

Where it fits

For teams already shipping generative work, three things make Krea Agents worth tracking. It is one of the first serious attempts to bring the agent pattern into visual production at scale. It plugs into existing agent clients through MCP. And it comes from a team that has shipped both an in-house foundation model and a product layer around dozens of external ones.

Use cases the team is highlighting include automating multi-step campaigns (generate variants, upscale, post to social), building a persistent brand style the agent can reuse across projects, and letting non-technical collaborators drive workflows that previously required someone comfortable with node graphs. If your current bottleneck is the graph-wiring step between idea and output, this is the layer aimed at removing it.