Skip to main content
June 29, 2026 Three MCP releases (0.21.0 → 0.22.0) and two panel releases (0.4.5, 0.4.6): the headline is parity with Comfy’s official cloud MCP — while staying local-first and far broader — plus run-to-node partial execution and a set of graph-navigation tools built for reading big workflows.

​
Comfy MCP parity — five one-call tools + a hosted connector

  • run_workflow_url — fetch a workflow from a shared / registry / raw-JSON URL, validate it (API or UI format, auto-converted), then load or run it. SSRF-hardened: DNS-resolved addresses are checked against private/loopback/metadata ranges, redirects are rejected, only bounded http/https fetches.
  • rerun_generation — re-enqueue the exact workflow behind a prior generation, with optional input overrides — reproducibility in one call.
  • generate_video — one-call LTX-2.3 text/image-to-video on the render-verified pack stack.
  • remove_background — one-call BiRefNet/RMBG cutout.
  • upscale_image — one-call model upscale.
  • Remote / hosted connector — token auth (Authorization: Bearer or X-API-Key, constant-time) on the Streamable-HTTP /mcp transport, plus a one-command public tunnel (npx -y comfyui-mcp --tunnel) that prints a paste-ready Claude Desktop Custom Connector URL + token. Binding /mcp to a non-loopback host without a token is now a hard error.
All five tools were live-verified on a real GPU (RTX 4090, ComfyUI 0.26.2) — which surfaced and fixed two graph bugs only visible against installed node schemas (generate_video missing required LTX/SaveVideo widgets; ComfyUI-RMBG KeyErroring on omitted “optional” widgets), now pinned by regression tests.

​
Run to node — partial-execution debugging

panel_run gains to_node_id: render only one output branch (the target output node plus everything upstream) via ComfyUI’s native partial execution, skipping every other branch — a fast, cheap way to preview or debug part of a big graph. The new debug-render skill encodes the method for renders that complete but look wrong: localize the first bad stage with run-to-node, preview-tap intermediates, fix, confirm. (Panel side ships in 0.4.5.)

​
Graph navigation — read a big live graph without dumping JSON

v0.22.0 + panel 0.4.6 add three tools built for an LLM to read a large canvas:
  • panel_graph_outline — a compact, dependency-ordered text map of the open graph (topo-sorted, each node with key widgets + ←/→ wiring, plus a groups index).
  • panel_find_nodes — search the live graph by type, title, port, widget name/value, output/subgraph/mode flags, or a free-text query, with a matched_on reason per hit.
  • panel_subgraph_group — wrap an existing group’s nodes into one toggleable subgraph node in a single step; panel_get_graph groups now report member node_ids.
Manual-edit awareness ships alongside: when you edit the canvas by hand between turns (bypass a node, change a widget, rewire), the agent’s next turn opens with a ”⟳ MANUAL CANVAS CHANGES” list and is told to treat it as ground truth over its memory of the graph.

​
Packs

  • wan-animate-ofm (0.21.1) — WAN Animate 2.2 video-to-video character animation, the “OFM hub” variant on the Kijai WanVideo stack. A personal pack: requires four private teskor-hub nodes and is static-validated only, with caveats documented.
  • artokun-flow now ships the subgraph-organized WAN Animate workflow (named subgraphs for loaders, preprocess, face lock, decode, SAM 3, upscale) — far easier to navigate, and sanitized for shipping.