AI assistants

Docket with local AI (LM Studio, Ollama, Jan, Open WebUI)

You can run your task assistant entirely on your own computer: a local model, a local MCP client, the local Docket connector and your own Vikunja. Nothing leaves your machine or network.

What you need

  1. Docket set up (the Windows installer, or a Docket server on your network).
  2. A model that's good at tool calling. Recent instruction-tuned models with tool support work best, for example Qwen 3 (8B or larger), Llama 3.1/3.3 (8B or larger), Mistral Small or gpt-oss. Very small models (under ~4B) often call tools with wrong arguments.
  3. An app that connects models to MCP servers. Ollama itself doesn't load MCP servers; pair it with one of the clients below.

#LM Studio

LM Studio is both the model runner and an MCP client.

  1. Load a model with tool use (the model list shows a hammer icon for tool-capable models).
  2. In the right sidebar open Program > Install > Edit mcp.json and add Docket:
JSON
{
  "mcpServers": {
    "docket": {
      "command": "C:\\Users\\you\\AppData\\Local\\Programs\\Docket\\mcp\\docket-mcp.exe"
    }
  }
}

On macOS/Linux add "env": { "DOCKET_VIKUNJA_URL": "http://localhost:3456", "DOCKET_VIKUNJA_TOKEN_FILE": "/home/you/.config/docket/vikunja-token" }. For a Docket server instead: "docket": { "url": "https://tasks.example.com/mcp", "headers": { "Authorization": "Bearer YOUR_CONNECTOR_KEY" } }.

  1. Save. In a chat, LM Studio asks before each tool call; approve it, or allow Docket's read-only tools permanently.

#Ollama

Use Ollama for the model and one of these as the MCP client:

  • Open WebUI (browser UI for Ollama): in Admin Settings > External Tools, add a server of type MCP (Streamable HTTP) with URL http://your-docket-server:8080/mcp (or https://tasks.example.com/mcp) and a Bearer key. Then turn the tool on in a chat. Open WebUI calls HTTP servers, so run the connector with --http (below).
  • Jan, Msty, Cherry Studio, 5ire and other desktop chat apps with MCP support: add a local (stdio) server with the docket-mcp command, like the LM Studio example, and pick an Ollama model.
  • Continue (VS Code / JetBrains): add Docket under mcpServers in Continue's config and choose an Ollama model.
  • Terminal: ollmcp connects Ollama models to MCP servers; give it the docket-mcp command as a server (see its README for the current flags).

#Run the connector over HTTP on your network

Some local clients only speak HTTP. Run the connector on the machine with Docket:

PowerShell
# Windows, with your Docket settings
$env:DOCKET_MCP_KEY = "a-long-random-key-of-24-or-more-characters"
$env:DOCKET_MCP_URLS = "http://127.0.0.1:8080"     # use 0.0.0.0:8080 to allow other devices on your network
& "$env:LOCALAPPDATA\Programs\Docket\mcp\docket-mcp.exe" --http

Then use http://127.0.0.1:8080/mcp with Authorization: Bearer <that key>. Plain HTTP is fine on one computer; across a network, put it behind HTTPS (Tailscale, Caddy) or use the Docker server.

#Tips for small models

  • Name the tool in your request: "Use list_tasks with view today." Small models follow explicit tool names well.
  • Keep the conversation short; start a new chat for a new job so the tool list and results fit the context window.
  • Give the model a system prompt such as:

    You manage the user's to-do list with the Docket tools. Task ids look like #123. Look tasks up before changing them, never guess ids, and confirm what you changed in one short sentence.

  • Raise the context length to at least 8k tokens; Docket's 26 tool descriptions take about 5k.
Edit this page on GitHub