ComfyUI Prompt Assistant: Cloud and Local LLM Prompting Inside the Node Graph
提示词小助手可以一键调用智谱、硅基流动、gemini、本地ollama、百度等大语言模型服务，实现提示词翻译、润色扩写、图片反推。支持提示词预设实现一键插入、历史提示词查找等功能。是一个全能型提示词插件。The Prompt Assistant enables one-click access to LLMs/VLMs for prompt translation, expansion, and image captioning. It also supports one-click preset insertion and historical prompt search.
At a glance
- What is it?
- Prompt Assistant is a ComfyUI extension that wires translation, expansion, image captioning and preset tags to a configurable set of LLM providers, including Ollama and any OpenAI-compatible endpoint. The design is provider-agnostic, but the project is a single-maintainer GPL-3.0 plugin whose release notes read as a running bug list.
- Who is it for?
- Adopt it if you already run ComfyUI and want prompt translation, expansion, captioning and reusable tag presets without leaving the graph, and you accept that the plugin is maintained by one author with a fast release cadence and a changelog dominated by fixes. Skip it if you need a stable API surface, if you cannot run ComfyUI 0.3.27 or newer, or if you want the extension to manage provider credentials for you.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Yes. The repository last received commits 144 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Prompt Assistant Actually Removes From the Workflow
ComfyUI prompts are text, and text is where the friction lives. A user writing in Chinese needs English tags for a model trained on English captions. A user with a reference image needs a caption before they can generate anything similar. A user who has settled on a favourite combination of quality tags and LoRA trigger words retypes it every session. Prompt Assistant targets all three by putting an assistant button on text widgets and image nodes, so translation, expansion and captioning happen where the prompt is edited rather than in a separate browser tab. The README describes the plugin as an all-in-one prompt tool, and the feature list matches that claim: prompt and Markdown node translation, node documentation translation, prompt optimisation, image and video captioning, tag favourites and history. The audience is ComfyUI users who already pay for or self-host an LLM and want it reachable from the canvas. It is not a model, not a sampler, and not a replacement for a prompt library. It is a client with a UI.
Provider Routing and the /v1 Split
The plugin talks to two families of backends. Cloud services (Zhipu, SiliconFlow, Gemini, Baidu, 302.AI and anything OpenAI-SDK compatible) go through an API configuration panel, and local models go through Ollama. The 2.0.6 notes describe the Ollama routing rule precisely: if base_url has no /v1 suffix the request uses Ollama's native API, and if it has /v1 the request uses the OpenAI-compatible API. That single rule explains a class of failures users hit when they copy a base URL from one tutorial and a model name from another. The same release note says the streaming filter logic was fixed because some models returned empty content and crashed the request. Configuration is per-task: expansion, translation and captioning can each select a different service, and multiple models can be registered as fallbacks. Version 2.0.0 moved user configuration to user/default/prompt-assistant so that reinstalling the extension does not wipe presets and tags. That is the right place for it, and it is worth checking after any upgrade that your tags survived.
Tags, Rules and Presets as CSV-Backed State
Version 2.0.0 changed the tag module to load CSV files, with multiple CSV sets switchable at runtime and a favourites mechanism on top. Rules got their own configuration window with categories and a setting for where a rule's button appears. In practice this means the reusable part of your prompting is data on disk, not code, so you can version it or share it. The history feature records per sentence, triggered when an input box loses focus, and supports undo and redo plus cross-node history viewing. That trigger choice matters: if you type and immediately queue the workflow without clicking away, the record may not be written. The README does not promise capture on every keystroke, and it should not be read that way. Preset insertion is one click from the tag panel, which is the feature most users will touch daily and the one least dependent on any LLM provider being reachable.
Image and Video Captioning, and Where It Breaks
Captioning sends an image (or images) to a vision-capable model and returns prompt text in Chinese or English, with styles such as natural language or tag lists. The video captioning node is labelled beta in the 2.0.0 notes, and the 2.0.6 fixes are instructive: some models, Qwen3.5-Plus among them, were misjudged as not supporting multi-image analysis, which raised an error. The fix infers an image limit from the model name and truncates automatically instead of aborting the task. That is a workaround for a real problem: there is no standard way to ask an arbitrary endpoint how many images it accepts, so the plugin guesses from the model string. If you use a model whose name the plugin does not recognise, expect either a conservative truncation or an error, and expect that behaviour to change between releases. This is the section of the plugin most exposed to provider quirks, and the changelog shows it.
Installing and the Version Floor
The README does not include a full installation command block in the material provided, so the exact clone path is not confirmed here. What is confirmed is the compatibility boundary: the V1.2.2 notes state that compatibility code was removed and ComfyUI versions below 0.3.27 are no longer supported. Version 2.0.0 states the frontend was rebuilt to support node2.0 mode, and 2.0.6 states the nodes were reworked for the ComfyUI V3 API standard. Configuration keys visible in the notes include base_url on the API panel, a direct-connection toggle moved into the settings interface in V1.2.2, a maximum token parameter added in V1.1.2, and per-node Ollama release options for freeing VRAM. Translation settings gained punctuation handling, automatic removal of extra spaces, removal of repeated dots and a preserve-newlines option in V1.2.1, plus a mixed-language rule parameter in V2.0.2. Those are the knobs to look for in the settings window. Anything beyond them is not documented in the supplied material.
The Cost of a Fast Release Cadence
Between 2025-06 and 2026-04 the project shipped at least a dozen tagged releases, and the changelog is dominated by repairs: Ollama 404s, Gemini 3 Pro requests failing, HTTP 502 through proxies, base_url trimming errors, tag and history modules breaking in 2.0.4 and fixed in the same entry, empty streaming responses crashing requests. Several entries describe regressions introduced by the previous release's refactor. This is normal for a plugin tracking a moving host application and a dozen provider APIs, but it has consequences. Pin a version you have verified rather than tracking main. Read the release notes before upgrading, because 2.0.x changed the tag storage format and the config location. And treat the node2.0 and V3 API work as ongoing: 2.0.6 is still fixing subgraph mounting in both Vue and LiteGraph modes. The licence is GPL-3.0, which matters if you plan to redistribute a modified copy or bundle it into a larger distributed product; that is a question for your own legal review, not something this article can settle.
How It Compares to a Generic OpenAI-Compatible Node Pack
ComfyUI already has node packs that expose an LLM call as a graph node, and several accept an OpenAI-compatible base URL. The difference is placement. A generic LLM node produces a STRING output that you wire into a text encode node, which is fine for automated pipelines but awkward when you are iterating on a prompt by hand. Prompt Assistant inverts that: the assistant lives on the text widget itself, so translation and expansion mutate the prompt in place, and the tag panel, history and undo operate on the same buffer. The trade-off is that you get less composability. If you want the LLM output to feed a conditional branch, a loop or a batch of prompts, a plain node with a string output is the better tool, and Prompt Assistant's per-widget buttons will not help you. Choose based on whether your prompting is interactive or programmatic.
Who Should Install It and What to Check First
Install it if you prompt by hand in ComfyUI, work across Chinese and English, and already have an API key or a local Ollama instance. The tag and preset features work without any provider configured, which makes the plugin useful even on an offline machine. Do not install it if you are on ComfyUI older than 0.3.27, if you need a frozen API that will not shift under you, or if your workflow depends on programmatic LLM calls that should be wired as graph edges. Before committing, confirm three things: your ComfyUI version, whether your Ollama base_url includes /v1 (this determines which API path is used), and that user/default/prompt-assistant exists after your first configuration change, since that directory is what survives a reinstall. If the tag CSV format from 2.0.0 does not match tags you exported from an earlier version, the migration path is described only in the release notes, so read them before you upgrade rather than after.
Editorial conclusion
Adopt it if you already run ComfyUI and want prompt translation, expansion, captioning and reusable tag presets without leaving the graph, and you accept that the plugin is maintained by one author with a fast release cadence and a changelog dominated by fixes. Skip it if you need a stable API surface, if you cannot run ComfyUI 0.3.27 or newer, or if you want the extension to manage provider credentials for you. Verify first that your ComfyUI version meets the floor stated in the V1.2.2 notes, that your Ollama endpoint answers on the path shape you intend to use (with or without /v1), and that your user data lands in user/default/prompt-assistant rather than the extension directory.
Community notes