@magpie-community/opencode-opendesign-auth
v0.1.13
Published
Open Design Cloud (Design Plan) in OpenCode and magpie: the account's device sign-in, its models, and the plan's 5-hour and 7-day allowance
Maintainers
Readme
@magpie-community/opencode-opendesign-auth
Open Design (open-design.ai) Cloud as a magpie / OpenCode provider: sign in with the Open Design account, use the models its Design Plan entitles the account to, and watch the plan's 5-hour and 7-day allowance and wallet balance where every other subscription's usage shows.
Works in magpie (which runs OpenCode provider plugins) and in OpenCode.
magpie plugin add @magpie-community/opencode-opendesign-auth
magpie plugin login opendesignIn OpenCode, opencode.json:
{ "plugin": ["@magpie-community/opencode-opendesign-auth"] }then opencode auth login.
What it signs in to
Open Design's cloud is two planes:
| Plane | Host | Key | Used for |
| --- | --- | --- | --- |
| Control | amr-api.open-design.ai | vela_ctrl_… | plan, wallet balance, the Design Plan's allowance windows, the workspace |
| Model | amr-link.open-design.ai/v1 | sk-amr-… | /models, /chat/completions, /responses, /messages |
A sign-in is the same device authorization the Open Design CLI
(vela login) runs:
- the plugin POSTs
/api/v1/cli/device-authorizationsand gets a code (4HYFNA) and an activation page (open-design.ai/cloud/cli/activate?…); - magpie opens that page; approving it says the CLI is signing in;
- the plugin polls it and receives both keys for that account, with the account's email and id.
Nothing else is needed, and nothing of the desktop app's own sign-in
(~/.amr/config.json, a different machine-issued pair of keys) is read or
written.
Sign-in methods
| Method | What it stores | Allowance |
| --- | --- | --- |
| Open Design account | the account's control + runtime keys | 5-hour and 7-day windows, wallet balance, plan |
| Open Design API key | a pasted sk-amr-… key | not readable: amr-api doesn't take API keys |
The account sign-in is the one that shows the allowance; the API key one is there for a machine that can't open the approval page.
Models
Nothing is listed per account at request time from anywhere but the vendor:
magpie calls the plugin's provider.models, which reads amr-link's live
/v1/models at that account's key on every listing, so what a signed-in
account is offered is what its plan entitles it to right now. Limits and
input modalities come from that same catalog; the quota multiplier (rate)
comes from rates.mjs, the plan's own table (below). What the catalog leaves
out of a model models.mjs knows is the seed's: its limits, and every input
kind the seed declares (#829: the catalog left image out of DeepSeek V4.1
Flash Eco, which takes images). A kind the catalog adds is added; a kind is
removed only by changing models.mjs.
Reasoning levels are not in the catalog, so models.mjs records the ones
measured against amr-link (#137): every model took reasoning_effort
low and high, and all but GPT-6 Luna thought at them, the ones the
catalog calls standard too. Those models reason at low and high;
GPT-6 Luna has no levels. A model the seed doesn't know reasons when its
vendor tier is advanced, with no levels until it is measured.
deepseek-v4.1-flash-eco is the same model on the low-priority, high-quota
schedule (the vendor's announcement: 模型效果完全无损), so models.mjs gives
it DeepSeek V4.1 Flash's limits, input kinds and levels; only the quota
multiplier differs (0.27× against 0.5×).
Quota multipliers
rate is the plan's deduction multiplier — what one dollar of model spend
takes out of the Design Plan allowance — not the catalog's cost word. The
table in rates.mjs is the vendor's published one,
模型额度与使用说明 (read
2026-11-06; Go, Plus, Pro and Max carry the same table):
| Model | Rate | | --- | --- | | DeepSeek V4.1 Flash | 0.5× | | DeepSeek V4.1 Flash Eco | 0.27× | | DeepSeek V4 Flash, DeepSeek V4 Flash Vision Exp, DeepSeek V4 Pro, GLM-5.3 Flash, Hy4 Preview, Kimi K2.7 Code, MiMo V2.6 Flash, MiMo V2.6 Pro, MiniMax M3.1 Flash | 1× | | Claude Haiku 5.5, GLM-5.3 Flash-X | 2× | | GPT-6 Luna | 4× |
The catalog's own cost words stay as the fallback for a model the table doesn't name yet; they disagree with the plan for most models (Claude Haiku 5.5 and GPT-6 Luna are "low" there but 2× and 4× here). DeepSeek V4.1 Flash's 0.25× "限时谷价" row is a time-limited promotional price, not a second model id, so the model keeps 0.5×.
models.mjs is the seed the provider has before anyone signs in. Refresh
it from the live catalog with:
magpie plugin login opendesign # once, so the key is on hand
bun refresh-models.mjs # or OPENDESIGN_KEY=sk-amr-… bun …The script keeps the names and measured levels already there — the catalog carries neither — and, as the plugin does, the input kinds, limits and words the catalog leaves out. It prints each kind it kept that way. It reads nothing else: no secret is written to the file or printed.
Requests go to https://amr-link.open-design.ai/v1, each model on the API
amr-link answers its thinking on (measured, plugins#141): chat completions
(@ai-sdk/openai-compatible) for most, the Responses API (@ai-sdk/openai)
for claude-haiku-5.5 and deepseek-v4-pro, and Anthropic's messages
(@ai-sdk/anthropic) for glm-5.3-flash. Every model reasons at low,
high and max.
That API is each model's default, not the only one: amr-link serves every
model on all three (keyblues, plugins#888). A request for a model on another
of them is signed as that API takes the key — x-api-key on /messages,
Authorization: Bearer on /chat/completions and /responses — whatever
the model is declared on, so a model set to another API in magpie's model
editor works there too.
Where the sign-in is kept
magpie's plugin-auth.json (mode 600), as an oauth account:
{ type: "oauth", access: <runtime key>, control: <control key>, accountId:
<email>, uid: <account id> }. Neither key expires, so there is nothing to
renew and refresh is empty.
Use
magpie plugin # opendesign (Open Design, 11 models) signed in
magpie provider test opendesign
magpie quota # opendesign · Design Plan Go · 5h 32% ↻3hReplacing a custom provider
The plugin signs in to the provider id opendesign. A custom provider
of that same id (Providers → the host and key by hand) takes the id first,
so the plugin's provider is then listed as opendesign-plugin, with no
models of its own. Delete the custom provider — magpie provider rm
opendesign — and the plugin's provider takes the plain id, so agent
configs pointing at opendesign/<model> keep working, now with the
allowance card.
The allowance in routing
The windows are gating windows (magpie's aside is left off), so magpie's
own routing reads them like any subscription's:
- a window magpie reads as used up makes the account count as used up, so
routing groups skip it and a direct call to
opendesign/<model>is refused with a 429 until that window renews; magpie provider account-cap opendesign <account> [percent|off]holds the account at a share of each window, and--window "5 hours" 70caps one window alone, which is how to leave headroom instead of spending the plan to the last credit.
Since Open Design falls back to the wallet balance once the Design Plan allowance is out, an exhausted window holds the account rather than spending the wallet. Set the account to spend its credits, or cap the windows, to choose which of the two magpie does.
