Drive a node image workflow from Claude Code with Genvas

Yes: Claude Code can drive a node-based image workflow in Genvas, because Genvas ships a local MCP server that exposes the board to the agent. This page shows how to connect Claude Code to Genvas, have it build a board, price it, run it and then edit that same board.

Dee, who makes Genvas, wrote this page.

On this page

Before you start

These steps were read from Genvas 1.3.9, dated 9 November 2026. Genvas downloads exist for macOS on Apple silicon, Windows 10 and 11 (x64) and Linux (.deb, x64), and you need no account. You also need a key for one of the providers listed under Settings, Providers. You pay that provider directly, and Genvas adds nothing on top.

Build, price and run a board from Claude Code

  1. Open Genvas. If it shows a licence screen, activate it. An agent can use the app only while it is open.
  2. Open Settings, Providers and add a key for the provider whose image models you want. The panel says: Keys stay on this computer. You pay each provider directly. Until you add a key, that provider shows No key.
  3. Open Settings, Spending and set a Daily limit. The app says agent runs stop only at the daily limit. The setting “Ask before a run costs more than” applies only to runs you start in the app, so it will not stop an agent. Once spending in the last 24 hours reaches the limit, paid agent calls are refused.
  4. Open Settings, General, Agents and click Connect on the Claude Code row. The button then reads Connected. Genvas adds a server named genvas under mcpServers in ~/.claude.json and installs a skill named genvas-prompting. The same button appears during onboarding.
  5. Start a new Claude Code session, as the app tells you to after connecting.
  6. Ask for a board in plain words, and name it. For example: “In Genvas, use a board called Teapot. Make one image of a glass teapot in soft window light. Tell me the price before you run anything.” The agent uses the board you name, or creates one.
  7. Watch the board fill in on the canvas. The agent reads the node types with describe_schema, lists models with list_models and builds the board with create_workflow. The board is a prompt node, then a media node, with a model node attached. Model settings live on the model node. The agent also keeps a note on the board with the goal and status.
  8. Read the price the agent reports. It prices the board with run_workflow and dryRun set to true, which spends nothing. A separate estimate tool returns the price of a single generation in US dollars. For OpenAI images, draft mode previews them cheaply.
  9. Tell the agent to go ahead. It runs the board with run_workflow and polls run_status until the run is done. Agent runs can use OpenAI Batch at half price. If you close or restart Genvas during a run, the run is marked as interrupted and the agent is told to run that node again.
  10. Check the outputs on the board. Results appear there in the app, and the agent checks them before it expands a batch. To also get files in a folder you choose, set Settings, Export, Export folder.
  11. Ask for a change in the same session and name the board again: “On Teapot, make the light warmer and run it again.” The agent uses the board you name, so the change lands on Teapot and not on a new board. It does a dry run before the paid run, as it did the first time.

If it does not work

Claude Code reports “Genvas isn’t running”. Open Genvas.app. If it shows a licence screen, activate it first, then repeat your request.

The Claude Code row reads Not connected. Click Connect under Settings, General, Agents, then start a new Claude Code session.

A provider shows No key or Key rejected. No key means nothing is saved for it under Settings, Providers. Key rejected means the provider refused the key you entered, so replace it with one the provider accepts.

Paid calls are refused. Your spending in the last 24 hours has reached the Daily limit under Settings, Spending. Raise the limit or wait.

Other ways to drive a node graph from Claude Code

The notes below follow each tool’s own page or repository as they stood between 19 June and 10 November 2026.

Wireflow is a hosted node editor with its own MCP server. In Claude Code you add it with claude mcp add --transport http wireflow https://www.wireflow.ai/api/mcp and sign in through /mcp.

Claude-Imagine is an MIT-licensed MCP server that sends workflows to a ComfyUI server running on your own GPU. It installs with npx claude-imagine@latest and needs Node.js 20 or newer, a running ComfyUI and at least one diffusion model.

A plain script is another route: Claude Code can write a Python script that sends a workflow JSON to a local ComfyUI instance at localhost:8188 and downloads the output image, as 8fs describes.

More guides & comparisons

Genvas: images, voiceover and music from Claude Code

Genvas is a desktop app that lets Claude Code make a video's images, clips, voiceover and music, then join them. Steps, price check and fixes.

Make images and video in Codex with Genvas

Genvas is a desktop app for AI images and video. Connect it to Codex over MCP and see the price of each run before it starts. You pay providers directly.

AI image generators that bill per use: 5 tools compared

Five image tools that bill per use, not monthly credits: Genvas (free beta), Leonardo.Ai, getimg.ai, MuApi and BlockRun, compared on price and keys.

Start your first board.

Add a key and make something in a minute.

Free during beta. Open source soon, under AGPL-3.0.

No subscriptionYour keys, your filesMac, Windows and Linux