Reported issues for comfyui-mcp

Pod holds 19 of 30 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to comfyui-mcp.

Most discussed

Request: enable Private Vulnerability Reporting (or share a security contact)

Hi — I'd like to report a security issue via coordinated (private) disclosure, but this repository doesn't have GitHub's Private Vulnerability Reporting enabled and I couldn't find a SECURITY.md or a security contact.

Could you either:

Enable Private Vulnerability Reporting — Settings → Security → "Private vulnerability reporting" → Enable, or share a private contact (email) I can send the details to?

Following coordinated-disclosure norms, I'm intentionally not including any deta

Read the thread · 2026-08-30 · closed · external user · 2 comments

[Bug] Antigravity backend redundantly prepends 29KB systemAppend to argv, overflowing Windows 32K limit and wasting tokens

Summary

In the Antigravity backend adapter (src/orchestrator/antigravity-backend.ts), runTurn() prepends the full ~29,000-character systemAppend preamble into the CLI prompt argument (-p "<text>") on Turn 1.

Because ensureMcpConfig() already merges the comfyui and panel MCP server definitions into ~/.gemini/config/mcp_config.json, the CLI discovers and registers all tools and schemas natively. Passing a 29KB plain-text description of those same tools over argv is redundant

Read the thread · 2026-08-25 · closed · external user · 2 comments

Support split ComfyUI code and data roots

Problem

Some valid ComfyUI deployments separate the checkout from mutable data:

main.py, .venv, and custom_nodes live in a code checkout
models, input, output, temp, and user live under a shared or mounted data root selected by launch flags / extra-model-path configuration

This shape is related to the data-only COMFYUI_PATH reported in #851 and the multi-root Desktop layouts discussed in #766. Today comfyui-mcp has one COMFYUI_PATH for both concerns, so there is no

Read the thread · 2026-08-19 · closed · outside contributor · 7 comments

Feature request: native Qwen Code CLI backend for Agent Panel

Summary

Please add Qwen Code CLI as a first-class Agent Panel backend alongside Claude, Codex, Gemini, Antigravity, and the OpenAI-compatible local providers.

Qwen Code now exposes an agent-oriented local daemon and ACP bridge, so treating it as a generic OpenAI-compatible endpoint would lose its session, streaming, permission, cancellation, and MCP behavior.

Current environment and verified behavior

qwen version: 0.21.8
qwen serve is available as an experimental loc

Read the thread · 2026-08-11 · closed · outside contributor · 3 comments

pi.dev is ..."different"

pi.dev comes with very almost nothing installed, just: read, write, edit, bash. Users write/install "extensions" to get more advanced capabilities.

To access MCP I had to install an extension: npm:pi-mcp-extension ~/.pi/agent/npm/node_modules/pi-mcp-extension

It seems that within sidepanel pi.dev cannot access MCP because that extension is not recognized? Or not accessible? I'm guessing...I don't know: <img width="437" height="102" alt="Image" src="https://github.com/user-attachments/assets

Read the thread · 2026-08-03 · closed · external user · 4 comments

pi.dev agent access

[Feature Suggestion]: pi.dev agent support

Description

Since pi.dev is open source and written in json i wonder if it be useful to create a comfyui mcp (or custom) extension. Perhaps some of the browser side logic could be moved to the agent. The agent would run as a separate process, and yet access the comfyui browser DOM for all the gritty internal details by calling the bridge (?) Benefits are the agent would have more real estate, ComfyUI browser real estate and avoid those pesky pag

Read the thread · 2026-07-22 · closed · external user · 4 comments

Codex turn can remain pending forever when app-server ignores turn/interrupt

Environment

comfyui-mcp 0.45.0 / current main (3077250)
Windows 11
Agent Panel using the Codex app-server backend (gpt-5.6-sol)

Symptom

A Codex turn can remain in the working state indefinitely while newly submitted panel messages accumulate as pending. Pressing Stop or using send-now does not recover it when the live app-server process stops answering turn/interrupt.

Root cause

CodexBackend.interrupt() awaits the turn/interrupt JSON-RPC request without a dead

Read the thread · 2026-07-22 · closed · outside contributor · 2 comments

Add Antigravity CLI (agy) backend for Google subscription users

Summary

Please add Google Antigravity CLI (agy) as a first-class Agent Panel backend for users who authenticate through their Google AI Pro/Ultra subscription.

The current gemini backend is no longer a practical subscription path for individual users. Google announced that, from June 18, 2026, Gemini CLI would stop serving requests for the Google AI Pro, Google AI Ultra, and free individual tiers; those tiers moved to Antigravity CLI. Gemini CLI remains relevant for enterprise/Google Cl

Read the thread · 2026-07-21 · closed · outside contributor · 3 comments

Most recent

[Bug] Default loopback HTTP MCP port (9180) still collides with Logitech G HUB on Windows (follow-up to #2030)

Summary

In #2030, the default WebSocket bridge port was moved to 9199 to resolve collisions with Logitech G HUB (lghub_updater.exe / lghub_agent.exe).

However, the loopback HTTP MCP server port (ports.panelMcp), which exposes panel_* live-graph tools to non-Claude CLI backends (codex, gemini, agy, qwen, grok), still defaults to 9180 in src/orchestrator/index.ts:1773.

Impact

On any Windows machine running Logitech G HUB software:

startPanelMcpHttpServer

Read the thread · 2026-08-25 · closed · external user · 2 comments

Lulu Ads — discovery + monetization for comfyui-mcp

Hey — 178 tools, 36 AI skills, local/LAN/VPS/Comfy Cloud support for ComfyUI — this is a seriously complete piece of infrastructure, impressive scope.

I'm building Lulu Ads (getlulu.dev), an opt-in listing + ad layer for MCP servers, useful as a discovery channel, with an optional disclosed-sponsorship layer if you'd ever want to monetize agent traffic (no code changes, fully opt-in).

Signup if interested (free, 70% rev share, 14-day POC where we cover API costs): https://getlulu.dev/publisher

Read the thread · 2026-08-24 · closed · external user · 1 comment

Skill description and body sizes past the documented Agent Skills limits

Six places in plugin/skills/ sit outside what the Agent Skills authoring guide asks for. One is past a stated maximum, the others are past recommended guidelines. I checked each one by opening the file, and the counts below are mine rather than a tool's summary.

Past the documented maximum

plugin/skills/report-bug/SKILL.md carries a description of 1524 characters. The frontmatter requirements give 1,02

Read the thread · 2026-08-15 · closed · external user · 1 comment

Codex run-completion previews can replay and compound into stalled sessions

Summary

Long-lived Codex conversations can become unusable because every completed panel_run automatically sends its output previews to the vision backend. Codex persists those localImage inputs as inline input_image data URLs, and its current compaction path retains and recopies them into later compacted.payload.replacement_history records.

This is an upstream Codex compaction bug, already tracked in openai/codex#33493 and related reports, but comfyui-mcp amplifies it through autom

Read the thread · 2026-08-13 · closed · outside contributor · 1 comment

Allow exact per-action policy for consolidated tools

Problem

#873 / #1383 added a fail-closed tool-level surface policy. That closes the compact call_tool bypass, but consolidated tools still make the policy all-or-nothing.

For example, keeping queue visible so an operator can allow list, status, and targeted cancel also exposes move, edit, and global clear. The same issue applies to other consolidated tools whose actions have different blast radii.

Proposal

Add an exact action allowlist:

COMFYUI_MCP_TOOL_ACTION_

[Read the thread](https://github.com/artokun/comfyui-mcp/issues/1452) · 2026-08-12 · closed · outside contributor · 1 comment

### UI-to-API conversion drops COMFY_AUTOGROW_V3 dotted inputs

## Summary

UI-to-API conversion drops linked entries of COMFY_AUTOGROW_V3 inputs. This affects core ComfyUI nodes such as ComfyMathExpression, whose repeated inputs are serialized as values.a, values.b, and so on without a separate values parent value.

## Minimal reproduction

1. Create a UI-format graph with PrimitiveInt connected to ComfyMathExpression input values.a.
2. Define ComfyMathExpression values in object_info as COMFY_AUTOGROW_V3.
3. Run convertUiToApi.

## Actual behavior

The con

[Read the thread](https://github.com/artokun/comfyui-mcp/issues/762) · 2026-08-03 · closed · outside contributor · 1 comment

### vendor files missing

I upgraded to comfyui 0.29.0.
I had some mcp confusions, so restarted that, but then my agent panel disappeared. 

FYI I also had some issues "updating" "ComfyUI Agent Panel"" via the "extensions" button. It said "Update" but it just hung on "Update". So I uninstalled and reinstalled and I believe that hung as well, but I just moved on. 

FYI I also ran "npm build" inside comfyui-agent-panel - some bad advice from a rogue agent I suppose. I got lot's of ENOENT... I mention it incase this overwro

[Read the thread](https://github.com/artokun/comfyui-mcp/issues/749) · 2026-08-03 · closed · external user · 1 comment

### Claude backend drops blocking SDK informational messages and reports an empty turn as success

## Summary

When a `UserPromptSubmit` hook blocks a prompt, the Claude Agent SDK emits a `system` / `informational` message containing the hook failure and marks it as preventing continuation. The Claude backend drops that event, then maps the following SDK `result` with subtype `success` to `ok: true`.

The panel therefore finishes the turn with no assistant text and no visible error. From the user side, Claude appears to start generating and then silently stops.

## Environment

- `comfyui-mcp

[Read the thread](https://github.com/artokun/comfyui-mcp/issues/740) · 2026-08-02 · closed · outside contributor · 0 comments

### PanelAgent stall watchdog duplicates the interrupted failure and releases the turn gate early

## Summary

`PanelAgent.onTurnStalled()` currently owns a visible stall warning, but then it releases the turn gate synchronously with `completeTurn()` and calls `backend.interrupt()` directly.

If the backend subsequently emits the expected terminal result

```ts
{ type: "result", ok: false, subtype: "interrupted" }

the result handler does not know that this was watchdog-initiated. The panel therefore shows a second, contradictory failure:

The agent stopped responding...
`That tur

Read the thread · 2026-08-02 · closed · outside contributor · 0 comments

list_mcp_tools is bloated40k-70k context token+security risk warning

I tested this mcp and it is unfortunately rather bloated. depending on setup i achieved two results, docker open webui native mcp causes a 40k context token load when this is used. mcp-app-bridge tool is double (70k), but thats because it loads it twice due to some ui detection)

the problem is that all the tools are packaged effectively as two massive tools. if the tools were separated by group or individual that would be great.

That would also permit this to be used in production, as it would

Read the thread · 2026-08-02 · closed · external user · 1 comment

"PLAN" box

The "PLAN" box takes up valuable realestate and sometimes seems to show activity when none is happening.

I loaded a workflow, then went out to download some models directly myself. Often this is necessary because things have moved or been updated so it's worth doing a little hunt. I grab the models and put them into subdirectories for organization. I'm not sure what "agent" is doing, but the "PLAN" box shows that a download is occurring, I asked it to stop, and it says it did. But the "PLAN" bo

Read the thread · 2026-07-30 · closed · external user · 1 comment

The remaining reports are on the project's issue tracker.