@d3ara1n/pi-provider-unisound

Unisound (云知声) MaaS provider for Pi Coding Agent — registers the U2 model family (plus plan-hosted GLM/Kimi) via the platform's OpenAI-compatible API.

Providers

Provider ID	Name	Billing	API Key Env
unisound	Unisound MaaS	Pay-as-you-go	$UNISOUND_API_KEY
unisound-plan	Unisound MaaS (Token Plan)	Subscription	$UNISOUND_PLAN_API_KEY

Both channels share the endpoint https://maas-api.unisound.com/v1. The keys are not interchangeable: pay-as-you-go keys come from Metered & Packs → API Key Management, Token Plan keys from Subscription Management.

Token Plan usage restriction: plan quota may only be used inside coding tools (U2Claw, OpenClaw, OpenCode and equivalents like pi). Using the plan key for non-coding automation or application backends is considered abuse and may get the key blocked.

Models

The catalog covers the full platform model set — both channels expose the same models and differ only in billing (permissions live on the API key). It is refreshed from the live API (GET /v1/models) when the cached snapshot is stale — at most once every 4 hours, matching pi's built-in catalog cadence — so newly published models appear without a plugin update:

• Network refresh: each registration fetches the live catalog with its own key during interactive startup when the persisted snapshot is older than 4 hours, and replaces the registered model list. pi update --models forces an immediate refresh.
• Offline restore: the last-seen catalog is persisted to ~/.pi/agent/models-store.json and restored on offline startups (PI_OFFLINE=1).
• Metadata: context_window / max_output come from the live catalog. Modalities, thinking behavior and pricing are not exposed by the endpoint and fall back to the shipped specs per model id; models new to the catalog get conservative family defaults (text-only input, on/off thinking).
• Fallback: the shipped spec table is used until the first successful network refresh (or while the API is unreachable).

Reasoning families (shipped specs)

Family	Thinking levels
U2 Flash	on/off (thinking.type)
U2	always on (cannot disable)
U2 Med	off / low / medium / high (reasoning_effort)
GLM-5.2, DeepSeek V4 family	off / high / max (reasoning_effort)
GLM-5.3 / 5.3 Flash	off / low / high / max (reasoning_effort)
Kimi K3	low / high / max, always on (reasoning_effort)
others	on/off (thinking.type) — levels unverified

u2-radimed (medical imaging) is no longer listed by the live catalog and was removed (2026-10); it re-registers automatically if it returns. Pay-as-you-go U2 pricing is in Pricing; every model on the Token Plan and unpublished-price third-party models cost 0.

Compatibility

• Thinking is on by default on the U2 models: u2-flash and u2-med can toggle it, u2 always reasons, and u2-radimed never does. u2-flash ignores reasoning_effort (re-verified live), so only on/off is offered for it.
• Reasoning effort applies to u2-med and the plan-hosted models: u2-med exposes off / low / medium / high, glm-5.2 exposes off / high / max, and kimi-k3 is always-on with low / high / max. Gateway alias levels (e.g. minimal → low on u2-med) are hidden.
• Image input: u2-flash is text-only; u2-med and u2-radimed accept images.
• Context overflow errors match pi's built-in patterns, so auto-compaction works. Declared context windows are conservative, so pi compacts before the gateway rejects an oversized request.
• Caching: the API reports cache-hit tokens, which pi shows as cache-read tokens in the cost estimate.
• Tool calls, streaming, and usage reporting follow the standard OpenAI shapes.

Pricing

pi tracks cost in USD. Unisound lists CNY prices, so cost fields are FX-converted estimates (~7.1 CNY/USD) for pi's status-bar display only — not real billing. List prices (input / cache-hit / output, CNY per million tokens): u2-flash 1/0.2/2 (limited-time promos such as the Sept–Oct 2026 free month excluded), u2 1/0.02/2, u2-med 8/2/28, u2-radimed 15/4/20. Token Plan models have subscription billing → cost 0. Cache-write pricing is not published.

Installation

pi install npm:@d3ara1n/pi-provider-unisound

Or add to ~/.pi/agent/settings.json:

{
  "extensions": [
    "/absolute/path/to/pi-extensions/packages/pi-provider-unisound"
  ]
}

Set your API key (pay-as-you-go):

export UNISOUND_API_KEY="sk-..."

Or for the Token Plan:

export UNISOUND_PLAN_API_KEY="sk-..."

Or use /login in pi to store them in ~/.pi/agent/auth.json:

{
  "unisound": { "type": "api_key", "key": "sk-..." },
  "unisound-plan": { "type": "api_key", "key": "sk-..." }
}

Then switch models with /model unisound/u2-flash or /model unisound-plan/kimi-k3.

Getting an API Key

1. Visit the Unisound MaaS platform
2. Pay-as-you-go: Metered & Packs → API Key Management
3. Token Plan: subscribe on the Token Plan page, then Subscription Management → dedicated API key

Note that model access is scoped per key — a key without permission for a model returns no model <id> permission.

Dependencies

None — this is a standalone provider with no pi-extension dependencies. It uses pi's built-in openai-completions streaming.

Usage Quota Reporting

Not yet implemented — Unisound MaaS does not currently expose a public quota or balance API. Token Plan credits are visible only in the web console (Subscription Management). When one becomes available, quota reporting will be added via @d3ara1n/pi-usage-block-core.