pi-provider-newapi

pi coding agent provider extension for self-hosted NewAPI AI gateways.

Registers a single newapi provider with dynamic model discovery, automatic cost calculation, and backend routing based on each model's enriched API metadata:

Recommended API	Endpoint
openai-completions, openai-responses	{baseUrl}/v1
anthropic-messages	{baseUrl}

中文文档

Installation

This is a pi coding agent provider extension. It has no runtime dependencies — only peer dependencies on pi itself.

pi install npm:pi-provider-newapi

Or install from git:

pi install git:github.com/ttimasdf/pi-provider-newapi

The extension is auto-discovered via the pi.extensions field in package.json — no extra configuration is needed after install.

Quick Start

Option A: Environment variables (fast path)

export NEWAPI_BASE_URL=https://ai.your-gateway.com
export NEWAPI_API_KEY=sk-your-api-key

The provider registers automatically on startup with discovered models.

Option B: Interactive /login (persist key)

export NEWAPI_BASE_URL=https://ai.your-gateway.com

pi> /login

Run /login in the pi session. pi will interactively prompt for the provider, and for the API key (sk-your-api-key). The key is saved to <agentDir>/auth.json by pi's built-in credential store. On next load, the extension reads it automatically.

Configuration

Base URL and model metadata are stored in <agentDir>/extensions/provider-newapi.json. The API key is managed separately by pi's built-in <agentDir>/auth.json (set via /login or NEWAPI_API_KEY env var).

<agentDir> defaults to:

OS	Path
Linux / macOS	~/.pi/agent
Windows	%USERPROFILE%\.pi\agent

{
  "baseUrl": "https://ai.your-gateway.com",
  "modelInfo": {
    "unknown-model-id": {
      "api": "anthropic-messages",
      "reasoning": false,
      "input": ["text"],
      "contextWindow": 128000,
      "maxTokens": 4096
    }
  }
}

On load, if NEWAPI_BASE_URL differs from stored, the base URL is updated. If NEWAPI_BASE_URL is not set, a warning is printed.

modelInfo entries are auto-generated for models not found in the built-in model database. Edit them to adjust api, reasoning, input types, contextWindow, or maxTokens. Optionally add thinkingLevelMap (e.g. { "xhigh": "max" }). When a previously unknown model later becomes known, the config entry is removed and upstream values take precedence, with a warning showing any differences between the config and upstream values.

How It Works

1. Config reconciliation — reads stored key and base URL; syncs with NEWAPI_BASE_URL env var
2. Ratio config fetch — fetches GET /api/ratio_config (no auth required) for NewAPI's model_ratio, completion_ratio, cache_ratio, and create_cache_ratio maps. This step is best-effort — if it fails, costs simply report as 0
3. Model discovery — fetches GET /v1/models from the gateway (requires API key)
4. Ratio matching — matches each model ID against ratio config keys using a three-step fallback: exact match → case-insensitive match → prefix match. This handles NewAPI's inconsistent naming (e.g. version tags, mixed casing)
5. Model enrichment — matches discovered models against built-in model data from multiple providers, in this priority order: deepseek, zai, google, anthropic, minimax, moonshotai, xiaomi, openai, vercel-ai-gateway. Models whose built-in API is not supported by NewAPI (openai-completions, openai-responses, or anthropic-messages) are skipped during lookup construction. Model IDs are normalized by stripping any provider/ prefix and converting to lowercase. Earlier providers win when the same model appears in more than one source. Enrichment populates api, contextWindow, maxTokens, reasoning, thinkingLevelMap, input types, and compatibility settings. During enrichment, supportsDeveloperRole is set to true for anthropic and openai sources and false for all others. Unknown models fall back to defaults (anthropic-messages, 128K context / 4096 max output tokens / text-only)
6. Cost calculation — converts from NewAPI quota to USD per million tokens. Based on NewAPI's formula Quota = (Input + Output × CompletionRate) × ModelRate × GroupRate with 1 USD = 500,000 quota:
   • cost.input = modelRatio × 2
   • cost.output = modelRatio × completionRatio × 2
   • cost.cacheRead = modelRatio × cacheRatio × 2
   • cost.cacheWrite = modelRatio × createCacheRatio × 2
7. Backend routing — each registered model carries its enriched api and endpoint, so pi routes through the appropriate built-in API handler instead of guessing from model ID prefixes
8. Model info templates — unknown models (not in built-in data) get a template added to provider-newapi.json under modelInfo for manual editing. When a previously unknown model later becomes known, the template is removed automatically and upstream values are used, with a warning logged showing any differences between config and upstream values

Graceful Fallback

If model discovery fails (network error, auth failure, etc.), the provider falls back to an unconfigured state:

• Registers a placeholder model newapi/unconfigured
• If an API key exists in auth.json, prompts user to run /reload then /model
• If no key exists, prompts user to run /login

This ensures pi never crashes on startup — it always presents a usable (if inert) provider.

Requirements

• pi coding agent
• NewAPI gateway
• Built-in model metadata from the supported upstream providers listed above
• For cost tracking: ExposeRatioEnabled must be true in gateway Settings → Operation → Ratio
• API key with access to models via the gateway

Without ratio_config

If the gateway has ExposeRatioEnabled set to false, the extension still works — all costs report as 0 (usage tracking disabled). Model discovery and routing are unaffected.

License

MIT © ttimasdf