Two open-source AI agents have appeared in the ComfyUI ecosystem, a node-based editor for image and video generation, that build workflows from text briefs. The first, agentY under the MIT license, works with VFX and image generation, builds and checks workflows itself, and controls professional 3D/VFX packages via MCP. The second, ComfyUI-Agent under Apache-2.0, is a beta local alternative to the cloud-based Comfy Agent from cloud.comfy.org in the form of a sidebar chat that sees the canvas. Both agents work with their own API keys or local models instead of credits for the cloud service and are already available for installation.

image
image

What happened

The agentY project (github.com/szprivate/agentY) has existed since February 21, 2026, and has accumulated 1,062 commits under the MIT license; the repository currently has 30 stars. It is an agent for VFX and image generation that works within the ComfyUI panel or via Claude Desktop: it builds workflows itself, searches the web for new workflows, compares the result with the brief, and redraws mistakes. Via MCP (a protocol for connecting external tools to LLMs), agentY controls Blender, Maya, Houdini, and Nuke, creates motion graphics via Remotion, works with Claude, GPT, Gemini, Qwen, DeepSeek, and local Ollama models, and supports Slack for communications. The second project, ComfyUI-Agent (github.com/CrazyDashTool/ComfyUI-Agent), was created on October 29, 2026, is distributed under Apache-2.0, and is in beta status: five commits, two stars. It is a sidebar chat within ComfyUI that sees the canvas, builds and validates workflows upon request, and works with its own API keys; it currently lacks response streaming.

Context

ComfyUI is a node editor for generative graphics where ready-made pipelines are traditionally assembled manually by connecting nodes; such assembly requires specific expertise and time. The ecosystem already has a cloud agent—Comfy Agent on cloud.comfy.org—operating on a credit-based payment model, meaning users pay for running tasks through third-party infrastructure. Both new projects offer a different framework: an LLM agent on top of the node editor that builds the graph from a brief and checks it itself; the core of this approach is a closed-loop self-correction cycle, where the result is compared with the original task before human acceptance. MCP in this scheme is a protocol for connecting external tools to LLMs, through which the agent gains control over third-party applications, from Blender to Nuke. The second important element is the bring-your-own-key approach: the agent connects to any chosen LLM, including inexpensive Qwen and DeepSeek or fully local Ollama, rather than a fixed cloud backend of the platform.

Why this matters for the industry

For the industry, this is a signal of a shift in the interface of the generative pipeline: instead of manual node assembly, the task is described in a brief, and the agent builds, validates, and reproduces the graph. This approach lowers the barrier to entry for VFX and motion studios, which no longer need a narrow ComfyUI specialist for every standard cycle, and agentY's mechanics—checkpoints, parallel agents "one per frame," memory, and project tags—look like a draft of a studio pipeline engine. Open-source MIT and Apache agents with bring-your-own-key are for the first time contrasted with the credit model of the cloud Comfy Agent: API keys and data are kept by the studio, and the cost of running is determined by the chosen LLM. The likely scenario for the coming months is convergence: open-source agents catch up to the cloud Comfy Agent in features, the first community evaluations appear, such as the success rate of building standard workflows, and pressure on the credit model of cloud.comfy.org grows. In the two-year perspective, if self-check cycles prove reliable, the interface of generative graphics will shift to declarative task description, and the role of manual pipeline assemblers will shift to designing agentic pipelines with a human in the loop.

Why this matters for users

Both projects can be installed today. agentY is cloned from the stable branch and installed via the install_agent.sh script or install_agent.ps1 on Windows or macOS 14+ on Apple Silicon; it requires an installed ComfyUI and any available LLM. ComfyUI-Agent is even simpler to install: git clone into the ComfyUI/custom_nodes directory without additional dependencies, and it has a Russian interface. The realistic benchmark is not "the agent replaces the artist," but reducing routine: rough workflow assembly from a description and re-checking the result. A reasonable tactic is a pilot on non-critical tasks where the cost of a redraw error is low. For those working in generative graphics, both tools provide ready-made ways to automate workflow assembly and result control while staying on their own keys and models.

What is currently unknown / limitations

Claims about the capabilities of both agents are not supported by articles, benchmarks, or evaluation methodologies—all information is taken from READMEs and a Telegram post, which is an interpretation of capabilities, not proven evidence. The key open question is the efficiency of the self-check cycle: how many iterations it takes to hit the brief, what it costs in tokens and time, and what the final hit rate is; without these figures, discussions about reliability remain a hypothesis. agentY, with 1,062 commits, has only 30 stars, meaning the external community has practically not validated the project: there are few bug reports and independent reproductions. ComfyUI-Agent, by engineering standards, is a prototype: five commits, two stars, beta status, no response streaming; any judgments about its reliability are premature.

Sources

Author

Look at AI, editorial team