Skills - NexLLM

Last edited: 4 Nov 2026 (Asia/Hong_Kong) Teach your agent how to get good results from NexLLM without memorizing model IDs. A skill is a small Markdown recipe (SKILL.md) the agent loads on demand; NexLLM MCP is what actually runs the work.

Available now. The official thin pack is public at NexLLMSupport/nexllm-skills tag v0.1.5 (MIT). Scope is text + image over MCP only — video, 3D, and audio outcome skills are not included yet, and this is not a Claude marketplace plugin.

Introduction

Agent Skills (the open SKILL.md folder format) are instructions, not a new API. Your agent matches a user request to a skill’s description, loads the recipe, then calls tools it already has. For NexLLM, those tools come from the MCP integration:

MCP tool	Role in a skill
`list_models`	Discover live model IDs for the connected key
`model_pricing`	Confirm cost before a paid run
`list_capabilities`	Optional — static supported/unsupported modalities (do not invent video/audio tools)
`model_metadata`	Optional — catalog fields for one model id from `list_models`
`generate_text`	Chat / completion through the gateway
`generate_image`	Image generation through the gateway

Spend lands on the wallet tied to the API key you pasted on connect.nexllm.ai. There is no skills-specific surcharge.

Prerequisites

A NexLLM account and API key at www.nexllm.ai
The NexLLM MCP server connected (https://mcp.nexllm.ai/mcp) with OAuth key-paste completed on connect.nexllm.ai
Wallet balance if the skill will call generate_text or generate_image
An agent that loads skills from disk (Claude Code / Claude with Skills, Cursor Desktop User skills, or any Agent Skills client)

Without MCP connected, a skill can only describe what to do — it cannot call NexLLM.

Install the official pack (v0.1.5)

git clone https://github.com/NexLLMSupport/nexllm-skills.git
cd nexllm-skills
git checkout v0.1.5

Pack contents:

Folder	Name	Outcome
`skills/nexllm-mcp/`	`nexllm-mcp`	Foundation: tools, discover → price → generate, refuse video/3D/audio
`skills/nexllm-models/`	`nexllm-models`	Discover / compare models and pricing
`skills/draft-with-nexllm/`	`draft-with-nexllm`	Draft or rewrite text via `generate_text`
`skills/product-hero-image/`	`product-hero-image`	Product / hero still via `generate_image`
`skills/product-photography/`	`product-photography`	Studio, lifestyle, and hero product stills
`skills/photoreal-stills/`	`photoreal-stills`	Realistic, camera-like still imagery
`skills/text-in-image/`	`text-in-image`	Posters / packaging / UI with exact on-image copy
`skills/logo-concepts/`	`logo-concepts`	Logo / icon concept rasters (not true SVG)
`skills/game-assets-2d/`	`game-assets-2d`	Sprites, stickers, and 2D asset stills
`skills/composite-scene/`	`composite-scene`	Prompt-only product + subject + backdrop (no mask API)
`skills/seedream-pro-layouts/`	`seedream-pro-layouts`	Layout / multi-element / poster-like stills; prefer Seedream 5.0 pro, then lite, then flash

Every generate skill requires list_models before hardcoding an ID. Image skills other than seedream-pro-layouts still suggest byteplus/Dola-Seedream-5.0-flash when that ID is present for the connected key. seedream-pro-layouts prefers byteplus/Dola-Seedream-5.0-pro, then a Seedream 5.0 lite id from the same live list, then flash. If you call byteplus/Dola-Seedream-5.0-lite or byteplus/ByteDance-Seedream-4.5, pass size: "2048x2048" — those IDs reject 1024x1024 with HTTP 400 plugin_request_invalid (size/aspect limits). The same eleven skill folders are also mirrored under .cursor/skills/ in the repo (for project / Cloud Agents discovery). The Cursor plugin layout continues to use skills/ plus .cursor-plugin/plugin.json (v0.1.5). Repo: https://github.com/NexLLMSupport/nexllm-skills · Install detail: docs/install.md

Add skills to your agent

Claude Code
Claude
Cursor
Other agents

Connect NexLLM MCP

Follow MCP integration (Claude Code tab) so list_models and the generate tools are available.

Clone the pack

git clone https://github.com/NexLLMSupport/nexllm-skills.git
cd nexllm-skills && git checkout v0.1.5

Copy into skills

Project-scoped:

mkdir -p .claude/skills
cp -R skills/nexllm-mcp skills/nexllm-models skills/draft-with-nexllm \
  skills/product-hero-image skills/product-photography skills/photoreal-stills \
  skills/text-in-image skills/logo-concepts skills/game-assets-2d \
  skills/composite-scene \
  skills/seedream-pro-layouts .claude/skills/

Or user-scoped for every project:

mkdir -p ~/.claude/skills
cp -R skills/* ~/.claude/skills/

Ask for the outcome

Start a session and ask in plain language (for example, “list cheap chat models on NexLLM”, “studio packshot of this bottle”, “poster with the exact headline Launch Week”, or “poster layout with three products”).

Connect the MCP connector

In Claude Desktop or claude.ai, add the NexLLM custom connector as described in MCP integration.

Open Skills

On a plan that supports Skills, go to Customize → Skills (wording may vary by client version).

Upload skill folders

Clone or download nexllm-skills @ v0.1.5, then upload / attach each folder under skills/ (each must contain SKILL.md). There is no Claude marketplace listing yet — use the files from the repo.

Ask for the outcome

In chat, describe the result you want. Claude should load the matching skill and call NexLLM MCP tools when needed.

Use Cursor Desktop and connect MCP

Cursor Desktop is the supported path for custom NexLLM skills. Connect https://mcp.nexllm.ai/mcp first (MCP integration — Cursor tab). Agents web can use NexLLM MCP tools after Custom MCP Connect, but the Agents web Skills picker does not load personal / custom skill packs today — use Desktop for skills.

Clone the pack (v0.1.5)

git clone https://github.com/NexLLMSupport/nexllm-skills.git
cd nexllm-skills && git checkout v0.1.5

Install User skills on Desktop

Copy each skill folder into your user skills directory (verified Desktop path):

mkdir -p ~/.cursor/skills
cp -R skills/nexllm-mcp skills/nexllm-models skills/draft-with-nexllm \
  skills/product-hero-image skills/product-photography skills/photoreal-stills \
  skills/text-in-image skills/logo-concepts skills/game-assets-2d \
  skills/composite-scene \
  skills/seedream-pro-layouts ~/.cursor/skills/

Project-scoped alternative (same layout under the project):

mkdir -p .cursor/skills
cp -R skills/nexllm-mcp skills/nexllm-models skills/draft-with-nexllm \
  skills/product-hero-image skills/product-photography skills/photoreal-stills \
  skills/text-in-image skills/logo-concepts skills/game-assets-2d \
  skills/composite-scene \
  skills/seedream-pro-layouts .cursor/skills/

Optional: in Cursor Settings → Agents, turn on Sync Skills for Cloud Agents. Sync alone does not make custom packs appear in the Agents web Skills picker — Desktop User skills remain the verified attach path.

Use a skill

In Desktop Agent (This Mac / local), open the skill picker (/ / Skills UI) and select a skill by name (for example /nexllm-mcp or /product-photography), or ask in plain language. Confirm the agent calls NexLLM MCP tools such as list_models.Optional alternate: Cursor Customize → import plugin/skills From GitHub using https://github.com/NexLLMSupport/nexllm-skills (plugin layout from skills/ + .cursor-plugin/). Prefer the Desktop ~/.cursor/skills/ copy when you need the verified User-skills path.

Pack tip: v0.1.5 adds seedream-pro-layouts (Seedream 5.0 pro → lite → flash after discovery) and keeps .cursor/skills/ mirrored for Cloud / project discovery. For day-to-day Desktop use, copying the eleven folders into ~/.cursor/skills/ is enough.

Connect MCP

Point the agent at https://mcp.nexllm.ai/mcp with OAuth. See MCP integration → Other clients.

Drop skills on disk

Most Agent Skills clients look under .agents/skills/ or ~/.agents/skills/. Confirm your agent’s docs, then:

mkdir -p .agents/skills
cp -R /path/to/nexllm-skills/skills/* .agents/skills/

Pair recipes with tools

Skills only describe the workflow. The agent still needs NexLLM MCP (or the Compatible APIs gateway) to execute.

How skills work with NexLLM

You ask → skill matches (name + description) → recipe loads →
agent calls NexLLM MCP tools → gateway runs → wallet billed

Design rules that keep skills correct as the catalog changes:

Name by outcome, not by a single model (product-hero-image, not gpt-image-2-only).
Discover before generate — call list_models (and optionally model_pricing) instead of hardcoding one ID forever.
Confirm the model is live for the connected key, then pass that exact id into generate_text or generate_image.
Stay within MCP capabilities — today that means text and image. Do not invent video or 3D steps the server cannot run.
Parse image results as JSON-in-text — generate_image returns structured text (URL or base64), not necessarily native image blocks.
Honest limits on image skills — logo-concepts produces raster concepts (not true SVG); composite-scene is prompt-only (no mask / layer API); seedream-pro-layouts is prompt-only multi-element layout (not Figma / editable layers).

What an agent does with one

What’s the cheapest chat-capable model I can use on NexLLM right now?

The agent should match nexllm-models (or nexllm-mcp), call list_models / model_pricing, and answer with real IDs for the connected key.

Make a clean white-background hero image for this sneaker.

The agent should match product-hero-image or product-photography, discover an image model, optionally price it, then call generate_image.

Poster with the exact headline: Launch Week.

The agent should match text-in-image, keep the copy verbatim in the prompt, and call generate_image.

Poster-like layout: headline band, three products, quiet footer.

The agent should match seedream-pro-layouts, discover an image model (prefer byteplus/Dola-Seedream-5.0-pro when listed), then call generate_image with regions named in the prompt.

Turn this packshot into a vertical launch video with voiceover.

Say clearly that NexLLM MCP does not expose video or TTS tools yet. Do not pretend a skill can call tools that are not on the server.

Verify

Ask	Expected skill	Expected tools
“Cheapest chat models on NexLLM?”	`nexllm-models`	`list_models`, maybe `model_pricing`
“Draft a polite reply from these notes on NexLLM”	`draft-with-nexllm`	`list_models` → `generate_text`
“White-background packshot of this bottle”	`product-hero-image` / `product-photography`	`list_models` → `generate_image`
“Photoreal street photo of a coffee cart”	`photoreal-stills`	`list_models` → `generate_image`
“Poster with exact text: Launch Week”	`text-in-image`	`list_models` → `generate_image`
“Logo concept for a coffee brand (raster OK)”	`logo-concepts`	`list_models` → `generate_image`
“Poster layout with three products and a headline”	`seedream-pro-layouts`	`list_models` → `generate_image`
“Make a vertical launch video with VO”	`nexllm-mcp` (refuse)	no fake tool calls

Write your own

Create a folder and SKILL.md with name + description frontmatter.
Point the workflow at NexLLM MCP tools only.
Prefer outcome names (restore-screenshot-text, draft-changelog) over model names.
Optionally add a references/ folder for longer prompt examples.
Test with MCP connected: ask the trigger phrase and confirm the agent calls the expected tools.

Starter patterns also live in the pack README.

Troubleshooting

The skill loads but nothing runs on NexLLM

The agent has instructions but no tools. Connect MCP, complete OAuth key-paste on connect.nexllm.ai, and confirm list_models works with a trivial ask.

Cursor Agents web does not list my NexLLM skills

Custom / personal skill packs are not attachable in the Cursor Agents web Skills picker today (even with Sync Skills for Cloud Agents ON). Install under ~/.cursor/skills/ on Cursor Desktop and use Desktop Agent — that is the verified path for nexllm-skills.

The agent hardcodes a model that 404s

Update the skill to require list_models before generate. Model IDs differ by channel group and key restrictions.

generate_image fails with plugin_request_invalid / size limits

Some Seedream IDs reject the common OpenAI default 1024x1024. For byteplus/Dola-Seedream-5.0-lite and byteplus/ByteDance-Seedream-4.5, use size: "2048x2048" (verified 2026-10-25). Prefer byteplus/Dola-Seedream-5.0-flash when you need a 1K square. The gateway returns HTTP 400 plugin_request_invalid when the size is outside that model’s pixel/aspect limits.

generate_image fails with a gateway error

Some image model families can return a gateway error depending on availability. Prefer a known-good image model from list_models — for example byteplus/Dola-Seedream-5.0-flash — then retry generate_image.If using a GPT Image id (nexllm/gpt-image-*), omit response_format in the tool/gateway body. If the client/tool defaults it to url, prefer Seedream flash or another known-good id until the tool omits the field for GPT Image.

The user asked for video / 3D / music

Decline or explain the gap. Current MCP tools are discovery, pricing, text, and image only. Do not fabricate tool names.

I expected a one-line marketplace install

Use the public GitHub pack (git clone + copy folders). A Claude /plugin marketplace entry is not published yet. On Cursor, Desktop ~/.cursor/skills/ copy is the supported install; optional From GitHub plugin import is available on some builds.

MCP integration — connect Claude, Cursor, and other clients
Compatible APIs — OpenAI-compatible gateway for apps
Image Generation API — gateway image endpoints behind generate_image
Model pricing — how wallet charges work
Vercel AI SDK — app-side AI SDK chat (not skills)

​Introduction

​Prerequisites

​Install the official pack (v0.1.5)

​Add skills to your agent

​How skills work with NexLLM

​What an agent does with one

​Verify

​Write your own

​Troubleshooting

​Related

Introduction

Prerequisites

Install the official pack (v0.1.5)

Add skills to your agent

How skills work with NexLLM

What an agent does with one

Verify

Write your own

Troubleshooting

Related