Magpie

Use your D.AI models — GPT-6, Claude Opus 5.5, Grok 4.7, DeepSeek V4.1, GLM 5.3, Gemini 3.8 — from any agent that speaks through magpie or OpenCode.

One ddsk_… API key is one account, and magpie keeps several.

Add it

magpie plugin add mdddj/d-ai-magpie       # from GitHub (the unofficial list installs this)
magpie plugin add @d-ai/magpie            # from npm, once published
magpie plugin login dai                   # asks for the API key
magpie provider test dai                  # one tiny call per API

Then pick a model anywhere magpie does: dai/claude-opus-5-5, dai/gpt-6-astra, dai/deepseek-v4.1-flash, …

The key comes from the D.AI console (/#/apikeys). Custom deployments can point at another host with magpie plugin options dai '{"baseUrl":"https://dai.example"}' or DAI_BASE_URL.

DAI_API_KEY=ddsk_… ./e2e.sh runs the same checks in a throwaway HOME, so it never touches the magpie you actually use. Add --live to make one real call on each API.

What it exposes

15 token-metered chat models, in two APIs:

API	Models	Path
OpenAI chat completions	deepseek-v4.1-flash, gemini-3.8-flash, glm-5.3-flash, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-6-astra, gpt-6-sol, grok-4.6, grok-4.7	/v1/chat/completions
Anthropic Messages	claude-fable-5, claude-opus-4-7, claude-opus-4-8, claude-opus-5, claude-opus-5-5	/v1/messages

The model list and the prices come from D.AI's public catalog on every start (GET /api/system/public/capability-docs), narrowed to the models that key may call (GET /v1/models), so a model you are not entitled to never shows up and a new model appears without a plugin update. A built-in list is used when D.AI cannot be reached while the provider is first configured.

Every model reports its context window, its output cap, its USD/1M-token prices and its thinking levels. Image input is marked for the families D.AI serves with vision (GPT-5/6, Claude, Gemini, Grok, DeepSeek, GLM).

Usage

magpie quota shows the account behind the key: the wallet balance, the tier and its cycle, how much of the balance this key alone may spend, and how many calls it made today. That comes from GET /api/ai/account, an endpoint authenticated by the ddsk_ key itself.

A D.AI key has no plan window — it spends the account's shared wallet — so the card reads as a balance, not a quota window. Two keys of one account show the same balance but their own reservations and call counts.

Notes

• Only chat models. D.AI's image, video, music and ASR models are separate capabilities with their own endpoints; agents here want a chat completion.
• The usage card needs GET /api/ai/account on the D.AI deployment. Against an older one (404) the plugin falls back to the key's call log and shows call counts instead of a balance.
• The plugin never rewrites request bodies. Thinking levels are offered as model variants, and the gateway passes reasoning_effort through.
• Requests go out with Authorization: Bearer <key> plus x-api-key, so either D.AI route works.