ComfyUI Open-Sources Its MCP Server for Local Use

What happened

ComfyUI open-sourced its MCP server on 18 August 2026, adding a local-transport mode that runs on the same machine as a locally installed ComfyUI. The release merges that local path with the cloud MCP that shipped in June 2026 into a single unified interface, so one agent session can route any task to whichever environment it fits.

Context

The Model Context Protocol is the layer that lets an AI coding agent — Claude, Cursor, Codex, or any MCP-compatible client — issue tool calls to a server. ComfyUI, a node-based GUI for composing and running diffusion and video-generation models on your own GPU, had only been reachable through that protocol via a cloud-hosted instance. The cloud MCP launched in June 2026. Comfy identifies the local MCP request as the number-one user ask that followed. This release answers that directly.

How it works

The MCP server now runs on the same machine as ComfyUI. It reads the user's GPU, enumerates every loaded node including custom ones, and lists every model file on the local disk, exposing all of that to the agent over the standard MCP protocol. Before a download, the agent checks the reported GPU and says whether a given model will fit on that card. It manages model files, starts the local ComfyUI instance, finds and installs custom-node packs, and gets a workflow to a runnable state. The blog's end-to-end example walks through checking the GPU, selecting a compatible MiniMax H3 version (an open-weight video model), pulling weights, finding a template, verifying nodes, and running the workflow with output to a local directory. Because the agent shares the local file system, it can also pull render passes from Blender, run a generation in ComfyUI, and write results back for DaVinci Resolve. The stated fastest setup path is a single prompt pointing an agent at the guide at docs.comfy.org/agent-tools/mcp.md.

Our read

The headline is "open-sourced the MCP server," and that is accurate, but the part that matters for a working studio is not the protocol layer. It is the cross-tool reach. An agent that reads a Blender scene file, queues a reference-to-video generation in ComfyUI, and drops the result into a DaVinci Resolve project folder — all on one machine, one conversation — removes the file-shuffling that makes a local pipeline feel like three separate jobs. No cloud round-trip, no manual export-import loop.

The GPU fit-check is the smaller but more immediate win. The most common failure in local video generation is downloading a model onto a card that cannot hold it, waiting for the transfer, and hitting an OOM crash. The agent reading the card and saying "this will not fit" before the download starts removes that. The sources do not state what thresholds the agent uses, so validate on your own hardware before trusting it for a production batch.

The cost structure is the quiet win. The blog states that the only running cost on local hardware is electricity. For a studio that overnight-batches scene variations, that is the difference between a per-generation cloud fee and a plug in the wall. Both transports live under one MCP server and one Comfy account, so the agent routes per task. No separate tooling, no separate credentials.

What the vendor is not saying: whether local-only use requires a Comfy account at all, or whether it is entirely free and account-free. That distinction matters for a studio that wants zero vendor dependency on the local path.

What this changes

On Monday: stop hand-configuring MCP JSON files and point an agent at the docs URL instead. For a studio already running ComfyUI locally, the immediate wins are concrete. The agent checks whether the next open-weight video model will fit your card before you start the download. It can find, install, and wire a custom node pack into an existing workflow, removing the browse-install-retest loop. If you composite in Blender or grade in DaVinci, the agent moves files between those tools and ComfyUI on the shared file system without manual export. Batch overnight runs cost electricity, not a per-generation fee. Reserve the cloud path for jobs the local card cannot handle.

License

The blog states the MCP server is "fully open source" but does not name a specific licence. The sources do not state whether it is Apache-2.0, MIT, or another permissive or restrictive licence. Check the LICENSE file in the repository at github.com/Comfy-Org/Comfy-mcp before building anything commercial on it.

Key takeaways

• ComfyUI's MCP server now runs locally alongside ComfyUI and is unified with the cloud path under a single interface, so one agent session covers both.
• The agent reads your GPU, full node list, and model directory, and checks model fit before a download, removing the most common local failure mode.
• Cross-tool operation across Blender, Houdini, DaVinci Resolve, and ComfyUI is possible because they share one local file system.
• Local generation costs only electricity; the cloud path is reserved for jobs the local card cannot handle.
• The specific licence is not stated in the blog post; confirm it in the repository before commercial use.

Sources

1. Open Sourcing Comfy MCP on Local — tier 1