uncomfymcp: a minimal ComfyUI MCP server for text-to-image chat workflows without exporting APIs
citrainmyhefeweizen · reddit · 2026-09-06
A developer open-sourced uncomfymcp, a minimal ComfyUI MCP server for the generate-review-tweak chat loop, driving text-to-image workflows directly from Claude and AnythingLLM.
Design goals:
- Run saved workflows without manual API-format export or graph edits (no inserted websocket save node)
- Switch workflows mid-chat (e.g. "with my Krea2 workflow, generate...")
- Attach full workflow metadata to generated images
- Inline chat preview of results
Limitations: no reference images/image-to-image, no workflow toggles, no async/batch, no auth; auto-patching relies on simple heuristics and may break complex workflows. GPL-3.0 due to comfy-cli dependency.
The author also surveys inline image support across chat clients: Claude Code and Linux VS Code work, Windows VS Code has a vanishing-image bug, AnythingLLM shows none, Claude collapses them.
More from coding & agent
- Running Qwen 3.8 27B Locally to Drive Blender via MCP and Build a 3D Llama — jacek2023 · 2026-09-06
- tokenbender's trick: ask the model for two prose narratives to detect long-task drift — tokenbender · 2026-09-06
- KV cache pressure tool shows vLLM's advertised 2M-token cache can retain 3M after fixes — t4a8945 · 2026-09-06
- Nous Research ships Hermes BOT MODE for building multi-agent AI teams — Teknium · 2026-09-06
- User says GPT-6-Astra built 33 analytics insights and 3 dashboards in 30 minutes — Scobleizer · 2026-09-06
- Student Open-Sources MCP Gateway That Cuts Token Use 50% With Vector-Indexed Tool Catalog — randomdick_18 · 2026-09-06