ComfyUI Copilot
Generate, debug, rewrite, and tune image workflows inside ComfyUI.
The README places some actions behind user interaction, such as clicking Accept to import a workflow, and indicates prompts before model downloads. It also discloses that Agent features require an API key and base URL. However, the supplied files define no permission model or least-privilege design and even recommend running ComfyUI with sudo. They do not fully explain where workflows, canvas contents, email addresses, keys, or prompts are sent, retained, or processed. Key entry and saving are documented, but storage, redaction, rotation, and log protection are not. Dependencies have no hashes, audit evidence, or demonstrated security-update process, and several are unbounded. Canvas changes, batch execution, and downloads are user-triggered but lack per-effect previews, confirmations, and safety boundaries. Recovery guidance is limited to clearing context or reinstalling, with no documented undo or workflow restoration. Attribution appears in the README, license, and contact address, but the reviewed object names ATH-MaaS while the files identify AIDC-AI in repository links, workflow ownership, and copyright; combined with unverified publisher identity, this supports only partial source attribution.
The README, pyproject, and publishing workflow consistently describe a ComfyUI custom node at a broad level, but README version 2.0.0 conflicts with pyproject version 2.0.28. There is also tension between the suspended API notice and claims that Agent capabilities remain fully available, plus a repository-owner mismatch between the object and supplied files. Dependencies and installation paths are listed, but many packages are unbounded, no lockfile is supplied, an external API is suspended, and Manager installation is explicitly described as bug-prone. Failure guidance covers update errors, missing models, long-context interruptions, and submitting logs, but mostly recommends reinstalling, clearing context, or posting screenshots rather than providing structured diagnostics and assured recovery.
The material thoroughly identifies beginners and ComfyUI workflow developers and covers generation, debugging, rewriting, parameter tuning, and node/model discovery scenarios. It states boundaries involving post-May-2025 models, context length, runnable-workflow prerequisites, and suspended services, but does not systematically define unsupported workflows, provider differences, or hazardous-operation boundaries. Buttons, selected-node context, and natural-language examples make triggers reasonably precise, although ambiguity handling and accidental-trigger controls are not described. Python 3.10+, Windows commands, ComfyUI Manager, OpenAI/LMStudio, and custom API key/base URL configuration provide good environment coverage; the sudo recommendation, acknowledged Manager instability, and absence of a compatibility matrix prevent full credit.
The README is usefully organized into introduction, features, installation, activation, configuration, contribution, contact, and license sections, with extensive visual examples, although the suspension notice and later feature descriptions are not fully reconciled. Git, Windows, and Manager installation paths are practical, but virtual environments, exact compatibility versions, and least-privilege setup are missing. Product naming is mostly stable within the files, while the ATH-MaaS versus AIDC-AI repository mismatch and differing version numbers reduce stability. Examples cover primary tasks and several troubleshooting cases, but there is no dedicated FAQ or systematic troubleshooting table. Limitations around new models, context length, Manager behavior, and service availability are candidly documented. The complete MIT text agrees with README and pyproject metadata, justifying full license credit. Versioning has only a major-update note and continuous-update advice, without a formal changelog, release history, or migration notes. A contact email, community channels, contribution path, and owner-gated publishing workflow establish a maintenance route, but the relationship among PublisherId yx9966, AIDC-AI, Alibaba branding, and the stated object owner is unexplained.
The README presents directly usable ComfyUI outcomes: one-click workflow import, debugging suggestions, workflow rewriting, batch parameter comparisons, and node/model recommendations. However, the supplied evidence has no implementation code, tests, or inspectable static output samples establishing schemas, validation, or repair quality. Full-lifecycle assistance offers plausible marginal value over manual workflow construction, but support is primarily promotional prose and demonstrations. Users can supply OpenAI or LMStudio configuration, yet token usage, request volume, batch-compute cost, data-transfer overhead, and resource limits are undocumented; service changes also add operational cost, so cost-benefit treatment is thin.
Major capability claims trace to README descriptions and media, while installation, version, licensing, and publishing claims have corresponding files. Still, no implementation code, tests, or inspectable output fixtures are included, so most behavior cannot be statically substantiated from this evidence set. README, pyproject, LICENSE, and the workflow partially corroborate name, license, and publishing mechanics, but versions, repository identity, and service status do not align cleanly. Some constraints are explicitly labeled as notes or service notices, yet claims such as high quality, precise issue identification, and complete availability are not separated from measurable evidence, leaving fact-versus-inference discipline weak.
- Do not follow the README's suggestion to run ComfyUI with sudo. Install as an unprivileged user in an isolated environment and review the node's filesystem, network, and execution permissions first.
- Before submitting workflows, canvas contents, email addresses, or keys, verify who operates the configured base URL and its retention and logging policies; the supplied source does not document these data flows.
- Dependencies are unlocked and several lack upper bounds. Produce a lockfile, run vulnerability scanning, and assess the security implications of older constraints such as sqlalchemy<2.0 and urllib3<2.0 before deployment.
- Back up ComfyUI workflows and configuration before enabling automatic rewriting, batch execution, or model downloads because no dependable undo or recovery procedure is documented.
- Verify the relationship among ATH-MaaS, AIDC-AI, Alibaba branding, PublisherId yx9966, and the registry publisher, and confirm the authentic maintenance and update channel for the assessed revision.
- The suspended-service notice conflicts with parts of the feature description. Confirm each currently available capability and the bring-your-own-model configuration requirements before relying on the product.
What does this agent do, and when should you use it?
ComfyUI Copilot is an intelligent assistant installed under ComfyUI's custom_nodes directory for building and iterating generative-image workflows. It launches from the left-side ComfyUI panel and provides conversational access to workflow generation, error analysis, workflow rewriting, node inspection, and model recommendations. Version 2.0 adds interfaces such as Debug, Model Download, and GenLab for checking canvas connections and parameters, locating missing models, and comparing batches of parameter combinations. Its chat and workflow-generation models can be configured separately with OpenAI or LMStudio, using a user-supplied API Key and Base URL. The deployment boundary is a local ComfyUI installation rather than a standalone assistant; the service notice also says the hosted API is suspended and that node information queries, job recommendations, and workflow generation will cease to be available, so adopters should confirm which target features remain usable with their own model endpoint.
After a user describes a desired result in the input box, Copilot can return three library workflows and one AI-generated workflow, each importable into ComfyUI with one click. Debug reads the workflow on the current canvas, analyzes parameter and connection errors, and returns repair or optimization suggestions; when it detects a missing model, it can prompt the user to download it, while Model Download supports keyword-based selection from recommended models. Users can request changes to the active canvas, including adding nodes, changing parameters, and restructuring workflow logic, with Clear Context available to limit long conversation histories. GenLab accepts parameter ranges, executes combinations in batches, and produces visual comparison results, provided the workflow already runs successfully. Documented supporting operations include node recommendations, node input/output and usage queries, downstream subgraph recommendations, and searches for base models or LoRAs.
- A new ComfyUI user who wants to describe an image task and receive a first workflow that can be imported directly.
- A workflow author troubleshooting a canvas that may contain invalid parameters, broken node connections, or missing models.
- A visual experimenter with a working workflow who wants GenLab to batch-test parameter combinations and compare the outputs.
- An experienced ComfyUI user who wants to add nodes, adjust parameters, or reorganize an existing workflow through conversational instructions.
- A user evaluating an unfamiliar ComfyUI node who needs its description, parameter definitions, inputs, outputs, and downstream suggestions.
- An image creator searching for a base model or LoRA that matches a written requirement.
What are this agent's strengths and limitations?
- Covers the workflow lifecycle from initial generation through debugging, rewriting, and parameter tuning directly within the ComfyUI canvas.
- Debug checks both parameters and graph connections and links missing-model detection to a download prompt.
- GenLab executes parameter combinations in batches and creates visual comparisons for systematic tuning of working workflows.
- The chat model and workflow-generation model can be configured separately, with explicit support for OpenAI and LMStudio.
- Its core operation is tied to ComfyUI and requires an existing installation, Python 3.10+, dependency setup, and access to the custom_nodes filesystem.
- The hosted API is suspended, so users must supply an API Key and Base URL; node information queries, job recommendations, and workflow generation are also listed as features that will cease to be available.
- The documentation warns that models released after May 2025, such as wan2.2, may not be understood correctly and can interrupt workflow rewriting.
- Complex rewrites carry substantial context, requiring users to invoke Clear Context regularly to reduce the risk of interruption.
- ComfyUI Manager installation is explicitly described as prone to bugs and may require removal, reinstallation, or log-based troubleshooting.
- The supplied repository identifier differs from the AIDC-AI repository used throughout the README, creating an upstream-tracking question for adopters.
How do you install or deploy this agent?
Python 3.10+ and an existing ComfyUI installation are required. The documented Git installation is:
cd ComfyUI/custom_nodes
git clone https://github.com/AIDC-AI/ComfyUI-Copilot
cd ComfyUI-Copilot
pip install -r requirements.txtFor the Windows embedded Python distribution, the documented dependency command is:
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-Copilot\requirements.txtAlternatively, open ComfyUI Manager, enter Custom Nodes Manager, search for ComfyUI-Copilot, and install it. The documentation warns that Manager installation is more prone to problems and recommends Git. Note that the supplied source repository is identified as ATH-MaaS/ComfyUI-Copilot, while the README's clone commands and repository links point to AIDC-AI/ComfyUI-Copilot.
How do you use this agent?
Start ComfyUI, locate the Copilot activation button on the left panel, and launch its service. In Settings, configure the chat model and workflow-generation model separately with OpenAI or LMStudio; because the hosted API has been suspended, enter your own API Key and Base URL for Agent capabilities. For a first run, type “I want a workflow for xxx.”, select one of the returned workflows, and click Accept to import it into the canvas. Use the Debug button at the upper right of the input box to inspect the current workflow, or type “Help me add xxx to the current canvas.” to request a rewrite. For parameter experiments, switch to GenLab, specify the ranges, and run the batch only after confirming that the workflow executes normally.