More local backends: LM Studio, llama.cpp, custom endpoints
- LM Studio as a first-class local backend (#160), with a full hands-off model/server lifecycle (#164) — the agent can load the model and start the server itself.
- llama.cpp (
llama-server) as a first-class local backend (#161). - Custom OpenAI-compatible endpoint as a first-class backend (#162) — point the panel at anything that speaks the API.
- Temperature pinned to 0 for these backends — nondeterministic empty finals after tool results are gone (#166).
Mobile building blocks: phone pairing + render mailbox
- On-demand phone pairing (#180) — a token-gated LAN/tunnel listener; nothing is exposed until you ask.
- Render mailbox (#182) — never lose a render while the phone is away; headless
(mobile/remote) tabs get renders delivered in-turn (#165) and
show_mediacan inline media bytes for clients with no ComfyUI reachability (#171).
gemma4 fine-tune Arena results
LLM Arena results for thegemma4-comfyui-mcp ladder: :e4b 14/20 (best local
model tested), :12b 13/20, and :e2b honestly flagged at 4/20 pending its v2
training fix. Leaderboard SVGs + model-sizing guidance updated across the docs and
the Ollama onboarding copy. Full story on the blog.
Workflow intelligence + node development
query_workflow(#179) — filter / traverse / aggregate over big workflows without loading the whole graph into context.- Graph-health findings in
validate_workflow/analyze_workflow(#175) — disconnected nodes, missing required inputs, duplicate model loads, orphaned branches, muted/bypassed nodes. - Node-dev tools (#173) — path-jailed read/search/write/patch plus per-pack git
for
custom_nodes; commit/push stays behindCOMFYUI_MCP_ALLOW_GIT_WRITES. get_comfyui_settings/set_comfyui_setting(#174) — read/write ComfyUI’s own user-settings store.calculate(#176) — safe batch math with variables and seeded RNG.panel_auto_layout(#177) — one-shot topological canvas auto-arrange, andpanel_connectauto-matches slots by type with full diagnostics on failure (#178).