Skip to main content
May 29, 2026 A big week — four releases (v0.6.0, v0.6.1, v0.7.0, v0.8.0) landed back to back. Here’s what changed for you.

New features

Declarative environment setup with apply_manifest

Spin up a reproducible ComfyUI environment from a single inline object or a JSON / YAML manifest: pip packages, custom nodes (registry IDs or git URLs with @ref pinning), and model downloads. Idempotent, with a per-item structured report so you can see exactly what was already installed, what just changed, and what failed. See Install & environment tools.

Custom-node authoring lifecycle

A complete write → test → publish loop, callable from your agent:
  • scaffold_custom_node — generate a Python node pack from a template, including .comfyignore, .gitignore, and (optional) a GitHub Actions workflow that publishes to the Comfy Registry on tag pushes.
  • verify_custom_node — restart ComfyUI with a bounded readiness wait and confirm the pack’s NODE_CLASS_MAPPINGS actually registered. Catches the silent “failed import → tool just doesn’t appear” trap.
  • publish_custom_node — publish to the Comfy Registry via comfy-cli. Your REGISTRY_ACCESS_TOKEN is never logged.
See Custom-node tools.

Cloud storage in and out

  • download_model now accepts s3:// and Azure Blob URLs in addition to HTTPS / Hugging Face, with a new s3 auth mode.
  • New upload_output pushes a generated image to S3, Azure, an HTTP endpoint, or Hugging Face and returns the URL(s).
See Models and Assets & images.

Server-side image conversion

New convert_image re-encodes a generated image (by asset_id or output-dir path) to PNG, JPEG, or WebP via sharp. Returns inline base64 plus an optional file write, and reports bytes saved — useful for trimming 4 MB PNGs down to ~300 KB WebPs before they leave your machine. See Assets & images.

Node discovery agent

A new comfy-researcher agent turns a problem statement (“I need a controllable upscaler that runs on 8 GB VRAM”) into ranked custom-node pack recommendations. It searches the Comfy Registry, evaluates candidates, and delegates deep dives to comfy-explorer. generate_node_skill now ships with a read-through cache keyed by source@version (COMFYUI_SKILL_CACHE_DIR), so repeat analyses are instant. Pass refresh to bypass. See Defaults, stats, and skills.

Media uploads

upload_video and upload_audio copy local video/audio files into ComfyUI’s input directory so workflows can reference them as inputs — mirroring upload_image (closes #12). See Assets & images.

Hosted documentation

The full tool reference now lives at comfyui-mcp.artokun.io/docs, schema-generated from the source so it stays in lockstep with the server.

Updates

Model downloads — auth, caching, deduplication

  • Per-request auth for gated and private hosts: bearer / basic / header / query. Secrets are validated, redacted from logs, and stripped on cross-origin redirects.
  • Content-addressed download cache with concurrent-download coalescing (no more double-downloading the same checkpoint) and optional LRU eviction. Configure via COMFYUI_DOWNLOAD_CACHE_DIR and COMFYUI_LRU_CACHE_SIZE_GB.
See Models.

Custom-node install — pin to a ref

install_custom_node now accepts a commit, branch, or tag — parsed from the URL (https://github.com/user/repo@v1.2.0), an explicit ref argument, or a repo@ref shorthand. Reproducible installs, no more “works on my machine” because main moved. See Custom-node tools.

Process supervision

  • Bounded startup readiness checks (COMFYUI_STARTUP_CHECK_INTERVAL_S / _MAX_TRIES) so a slow boot doesn’t hang the agent indefinitely.
  • Opt-in crash auto-restart with a bounded budget (COMFYUI_ALWAYS_RESTART, COMFYUI_RESTART_MAX_ATTEMPTS, COMFYUI_RESTART_WINDOW_S) — recover from transient OOMs without spinning forever on a real bug.
See Process control.

Better job error reporting

get_job_status and completion notifications now surface ComfyUI execution errors with the failing node ID and type, exception type/message, a truncated traceback, current_inputs, and an OOM flag — plus optional per-node and total execution timing. Fully additive. See Workflow execution.

Plugin skills

Two new skills bundled for AI assistants:
  • comfyui-frontend-extensions — authoring against the v2 @comfyorg/extension-api, including a v1 → v2 migration guide.
  • comfyui-node-registry — node authoring and publishing patterns.

Fixes

  • generate_node_skill’s cache now resolves the current pack version before lookup, so you never get stale docs after a pack updates. Writes are atomic (temp file + rename with content-hash check).
  • Spawned ComfyUI child processes now have error listeners, so a missing or unexecutable ComfyUI binary can’t crash the MCP server itself.
  • install_custom_node git refs are validated and passed via git checkout --end-of-options <ref>, closing an argv-option-injection vector.
  • Native dependency build scripts (better-sqlite3, sharp) are now permitted under pnpm 10’s stricter default, so fresh installs no longer silently skip them.

Experimental

An embedded-agent backend POC is available behind COMFYUI_MCP_AGENT_POC: a cloudflared quick-tunnel helper plus an AI SDK /api/chat endpoint with bearer auth, a request body cap, and a server-side model allowlist. Not part of default startup. Feedback welcome.
Questions, bugs, or wishlist items? Drop an issue at artokun/comfyui-mcp.