How do you use AnimatedWebs with a local AI model on Lemonade?

What Lemonade Server is

Lemonade is an open source local AI server. Its README describes it as a way to get the capabilities of cloud APIs on your own NPU and GPU, free and private, and says the Lemonade Server flavor is a service that apps can connect to through standard OpenAI, Anthropic and Ollama APIs. Models run on CPU, GPU or NPU, and the project is licensed under Apache-2.0.

It installs on Windows, Linux (Arch, Debian, Fedora, Ubuntu and Snap), macOS and Docker. It is a community project with optimizations from AMD engineers for Ryzen AI, Radeon and Strix Halo PCs. Lemonade and AMD are names of their owners, and this guide is not made by or affiliated with either.

Lemonade does not search or build websites by itself. It serves models. To get templates and build prompts into a conversation with a local model you need a client that supports MCP, which is the open standard AI tools use to call outside tools. That client talks to Lemonade for the model and to the AnimatedWebs MCP server for the tools.

One point of confusion: Lemonade documents its own MCP gateway at POST /mcp, but that makes Lemonade an MCP server so other tools can call your local models. That is a server, and the gateway supports tools only, with no resources or prompts. Lemonade's own app is a different matter: its chat docs say the Tools menu can choose tools from available external MCP servers, but its operations page says the current app reports external MCP administration as unavailable while a local client host is being introduced. So there is no documented way yet to add AnimatedWebs inside Lemonade's own app, and this guide uses a separate MCP client, as below.

Install and start Lemonade Server

1. 1Install for your system from the Lemonade install guide: the Windows MSI, the macOS .pkg from the latest release, a Linux package or Snap, or Docker.
2. 2Let the server start. The Lemonade Continue guide says the server starts automatically after installation, and that lemonade status verifies it.
3. 3Check the address. The documented default is http://localhost:13305, and the OpenAI-compatible endpoints sit under /v1, for example http://localhost:13305/v1/chat/completions. The Open WebUI guide uses http://localhost:13305/api/v1, and Lemonade's API overview lists both /v1/ and /api/v1/ route prefixes. If you change the port, change it in every client below.
4. 4Download a model with the CLI, for example lemonade pull Qwen3-Coder-30B-A3B-Instruct-GGUF, which is the example in Lemonade's Continue guide. The Model Manager at http://localhost:13305 does the same from a browser.
5. 5Load it before you start working. The Lemonade guides suggest pre-loading, with the tray icon or lemonade load <model-name>.

lemonade status
lemonade pull Qwen3-Coder-30B-A3B-Instruct-GGUF
lemonade load Qwen3-Coder-30B-A3B-Instruct-GGUF

If your Lemonade server sets an API key (the LEMONADE_API_KEY variable), clients send it as a Bearer token. The Open WebUI guide says the key field is unused by default but required by the form, and suggests a single dash.

Clients that can use a Lemonade model and the AnimatedWebs MCP server

Each client below is documented to do both jobs: talk to an OpenAI-compatible or Anthropic-compatible local server, and connect to a remote MCP server over Streamable HTTP. The AnimatedWebs MCP address is the URL in each snippet below. Sign-in is a browser sign-in (OAuth) where the snippet relies on the client's own sign-in, and an API key sent as an Authorization header where the snippet includes one. Every account has a key on the developer page of the dashboard, free accounts included, and a free account opens the free template prompts.

Lemonade's own docs cover the first three. For Cline and goose, Lemonade has no page, so the model side relies on those tools' own documentation for OpenAI-compatible servers and on Lemonade's statement that its API is OpenAI-compatible. We have not run any of these combinations. If a step fails, the troubleshooting pages linked from each MCP setup guide are the next stop.

Lemonade's built-in chat is not listed here because its operations page says adding external MCP servers is currently unavailable in the app. Check that page again, because it describes this as a work in progress.

Claude Code

Lemonade's Claude Code page says that lemonade launch claude configures Claude Code to use your local server through Lemonade's Anthropic-compatible API, and that --model skips the menu and starts that model. Add AnimatedWebs once with the command below, then run /mcp inside Claude Code to sign in. Lemonade's page does not say whether its launcher changes where Claude Code reads its MCP settings, so if animatedwebs is missing from /mcp in the launched session, run claude mcp list to see what that session can see.

# 1. Start Claude Code on a local model (Lemonade docs)
lemonade launch claude --model Qwen3.5-35B-A3B-GGUF

# 2. In another terminal, add AnimatedWebs once
claude mcp add animatedwebs --scope user --transport http https://mcp.animatedwebs.com

# 3. In Claude Code, run /mcp and choose animatedwebs to sign in

Continue

In the Continue sidebar open the model dropdown, choose Add Chat Model, and pick Lemonade Server. Lemonade's guide says Continue sets the default address to http://localhost:13305 and detects the server on localhost, and you then pick the model you downloaded. Next add the MCP server to the mcpServers block of config.yaml and switch the chat to agent mode, because Continue's docs say MCP tools only work there.

mcpServers:
  - name: AnimatedWebs
    type: streamable-http
    url: https://mcp.animatedwebs.com
    requestOptions:
      headers:
        Authorization: Bearer aw_your_key

Open WebUI

Follow Lemonade's Open WebUI page to add Lemonade as a connection with the URL http://localhost:13305/api/v1. Then, as an admin, go to Settings, Admin, Integrations, and under External Tool Servers click Add Connection. Open WebUI's MCP page says native MCP needs version 0.6.31 or later, that only administrators can add MCP servers, and that Type must be MCP (Streamable HTTP), not OpenAPI.

Type:  MCP (Streamable HTTP)
URL:   https://mcp.animatedwebs.com
Auth:  Bearer
Key:   aw_your_key

Open WebUI's MCP page also warns that MCP connects the tools but does not improve the model, so a weak model may still ignore them. If that happens, try a stronger model with reliable native tool calling.

Cline

Cline's docs say to set the API Provider to OpenAI Compatible and fill in a base URL, an API key and a model ID. Use the Lemonade address http://localhost:13305/v1, a placeholder key such as the dash that Lemonade's Open WebUI page suggests for its unused key field, and the name of your downloaded model (lemonade list shows model names). Then add the MCP server in cline_mcp_settings.json, with the type set explicitly because Cline defaults to the older sse transport.

{
  "mcpServers": {
    "animatedwebs": {
      "type": "streamableHttp",
      "url": "https://mcp.animatedwebs.com",
      "headers": { "Authorization": "Bearer aw_your_key" },
      "disabled": false
    }
  }
}

goose

goose's provider page says custom providers can use OpenAI, Anthropic or Ollama compatible API formats. In goose Desktop go to Settings, Models, Configure providers and click Add Custom Provider at the bottom of the window, or choose Custom Providers in goose configure. Pick the OpenAI Compatible type, then enter the Lemonade address as the API URL and your model name. goose's docs say a provider that needs no key has a checkbox for that in Desktop. Then add AnimatedWebs as a Remote Extension (Streamable HTTP). goose's docs say remote extensions that need OAuth usually get a client ID automatically, and this server supports Dynamic Client Registration.

goose configure
# Custom Providers: add an OpenAI-compatible provider
#   use http://localhost:13305/v1 as the address of the OpenAI-compatible API
# Add Extension > Remote Extension (Streamable HTTP)
#   name it animatedwebs and enter the URL below

goose session --with-streamable-http-extension "https://mcp.animatedwebs.com"

Which local models handle tool calling

Lemonade does not publish a ranked list of models for tool calling, so this section sticks to what its guides and the clients say. In its image generation section, Lemonade's Open WebUI page says to use a model that supports tool calling and suggests trying models larger than 30B, such as GPT-OSS-120B, GLM-4.7-Flash or Qwen-3-Next-80B-A3B. Its Continue page says a code-tuned model with at least 20B parameters is required for best results, and lists a PC with a discrete GPU of 16 GB of video memory or more, or a Strix Halo PC with 64 GB of system memory or more, as recommended hardware.

Open WebUI's docs say the connection can work while a model skips tools or sends wrong arguments, and goose's docs say it relies heavily on tool calling. So start with the largest tool-calling model your machine runs comfortably, and try a few. A model that can drive the three calls in the next section is enough for AnimatedWebs.

Context matters too. Lemonade's guides show lemonade load <model-name> --ctx-size 8192 as an example for large inputs. A full build prompt plus your conversation needs room, so raise the context size as far as your memory allows.

Build a template with a local model

Once the client shows the animatedwebs tools, ask for the work in small steps. A local model follows a short, specific request better than one long one.

1. 1Find a design. Ask: use animatedwebs to search templates for a dental clinic website and list the three best matches with their live demo links.
2. 2Pick one and look at its demo in your browser first, so you know what you are building.
3. 3Fetch its prompt. Ask: call get_template_prompt for the template you picked, using its slug from the demo address, and build it in this folder. The tool returns the build prompt for that template.
4. 4Build in passes. Ask for the hero first, then each section, and check the result in the browser after each pass rather than at the end.

The same flow works for backgrounds, UI components and style references through the other AnimatedWebs tools, all listed on the MCP page. Everything the model writes stays on your machine; only the tool calls to AnimatedWebs and the demo pages you open in a browser use the internet.

Limits to expect

• Local models are slower than hosted ones on most machines, and a long prompt can take noticeably longer to read before the first word appears.
• They may follow a long build prompt less closely. If a section is missed, paste the missing part of the prompt back and ask for only that section.
• Pro prompts still need sign-in. A free account opens every free template prompt, and a Pro prompt needs a Pro or Lifetime plan, or a Prompt Pack credit, on the account behind your sign-in or key. Local inference does not change that.
• Tool calling depends on the model. If nothing calls the tools, switch to a larger model before you change any settings.
• We could not run Lemonade for this guide, so none of the steps above were tested by us. Each one is taken from the linked official docs and from our own MCP setup pages. Tell us if a step needs a fix, using the contact page.

Sources

Facts about Lemonade and the clients were read on October 31, 2026 on the pages listed under Facts and sources at the bottom of this page. The date shown in the page header and in that box is the site-wide review date for all guides. Docs change often, so check the linked page if a setting looks different. AnimatedWebs facts come from our own MCP setup pages for Claude Code, Continue, Cline and goose.