Docs · Guide

WorkBuddy on any model

WorkBuddy is Tencent's CodeBuddy desktop app. It reads your own models from ~/.workbuddy/models.json (or $WORKBUDDY_CONFIG_DIR when set), as OpenAI chat-completions models. magpie writes its models into that file, so WorkBuddy's model picker lists DeepSeek, GLM, Kimi, Qwen, Claude, a ChatGPT subscription or any other model in magpie.

Set up WorkBuddy

  1. Add a provider in magpieAdd an API key on the Providers page, or sign in to a subscription.
  2. Set WorkBuddy's provider to magpie on the Agents pageWorkBuddy's picker then lists the models magpie picked for it.
  3. Pick the model per task in WorkBuddyThe model is picked for each task in WorkBuddy's window; magpie's show as custom-local:<id>. If they don't show, restart WorkBuddy; magpie's Agents page offers a restart.

Why a restart: WorkBuddy 5.6 watches models.json only when the file existed at start, and it watches the file itself, so a rewrite can go unseen until it restarts.

models.json

The file is either a list of models or {"models":[…],"availableModels":[…]} as CodeBuddy Code's CLI writes it; magpie reads and writes both. WorkBuddy's own local models are entries in the same list (local: true), and magpie leaves them alone. Each of magpie's models points at the gateway's /v1/chat/completions, with tool calling on and magpie's key for WorkBuddy:

Terminal
// ~/.workbuddy/models.json (one of magpie's entries)
{
  "id": "deepseek/deepseek-v4-pro",
  "name": "…",
  "vendor": "magpie",
  "apiKey": "",
  "url": "http://127.0.0.1:3425/v1/chat/completions",
  "maxInputTokens": …,
  "maxOutputTokens": …,
  "contextWindow": { "supportedLengths": [200000, …], "defaultLength": … },
  "supportsToolCall": true,
  "supportsImages": …,
  "supportsReasoning": true,
  "reasoning": { "supportedEfforts": […], "canDisableThinking": …, "defaultEffort": "…" }
}

WorkBuddy sends requests to url with reasoning_effort, under the apiKey as a Bearer token. Its requests say CLI/<version> WorkBuddy/<version>, so magpie tells them apart by their key. An entry you edited in WorkBuddy is saved under vendor Custom, and magpie still recognises it by its key.

Context and reply length

WorkBuddy offers a choice of context in its model menu only when contextWindow lists two sizes or more. magpie adds a smaller 200K beside the whole window, and defaultLength is the whole window. Without maxOutputTokens WorkBuddy caps every reply at its own default, so magpie writes it.

What stays yours

WorkBuddy rewrites models.json itself a few seconds after every start: it purges local models this computer can't run, and may encrypt the apiKey. magpie doesn't take that rewrite as a change. The other entries and keys stay as WorkBuddy wrote them:

WorkBuddy AI

WorkBuddy AI, the international build (www.workbuddy.ai, com.workbuddy.workbuddy-ai), is the same app with its data in ~/.workbuddy-ai. magpie wires it as an agent of its own, under its own key. On a Mac magpie finds it by its app, not by that folder. WorkBuddy in WSL works too, with the gateway's address as the distro reaches it.

Switching back

When you switch WorkBuddy off in magpie, magpie takes its models out of models.json and leaves every other entry as it was.

FAQ

Can WorkBuddy use DeepSeek or GLM?

Yes. magpie writes its models into ~/.workbuddy/models.json; pick one starting with custom-local in WorkBuddy's model picker.

Why don't magpie's models show in WorkBuddy?

WorkBuddy may not have seen the rewrite. Restart it; magpie's Agents page offers a restart.

Does magpie change my WorkBuddy settings?

No. Models you turned off, models pointed at a NAS, reply limits you typed and your default effort all stay.

Does WorkBuddy AI, the international build, work?

Yes. Its data is in ~/.workbuddy-ai, and magpie wires it as an agent of its own.

More agents on the same models: VS Code on any model, CodeBuddy on any model, Claude Science on any model and Pencil on any model. Something fails? See Fixing common errors, or ask on Discord.