Lightning AI provider for pi

A pi extension that exposes Lightning AI's OpenAI-compatible model gateway as the lai provider.

Features

• Discovers Lightning AI's official model catalog at startup.
• Filters user-created agent records and duplicate IDs from the public catalog.
• Imports current context windows, modalities, output limits, and pricing.
• Uses the Responses API for GPT-5.4+ and GPT-6 tool calls with reasoning.
• Reports token usage and calculates request costs in pi.
• Supports pi's /login credential flow and LIGHTNING_API_KEY.
• Keeps a static seven-model fallback for offline startup.
• Does not replace or interfere with pi's official providers.

Install

Requires Pi 0.99.2 or newer and Node.js 22.19 or newer.

From npm:

pi install npm:pi-lightning-ai-provider

Or directly from GitHub:

pi install git:github.com/mindthemath/pi-extensions-lai

Review extensions before installing them: pi extensions execute with your full user permissions.

Authenticate

Start pi, run /login, choose Lightning AI, and paste your API key. Pi stores it in its normal credential store.

Alternatively, set the environment variable before starting pi:

export LIGHTNING_API_KEY='...'
pi

The extension never reads project .env files and contains no API keys.

Select models

Run /model and filter by provider lai or by the LAI: display-name prefix. The model IDs remain Lightning AI's exact IDs, while the separate provider ID prevents collisions with direct OpenAI, Anthropic, or Google models.

The extension currently recognizes official catalog records whose server-side provider and ID namespace agree:

• lightning-ai → lightning-ai/*
• Anthropic → anthropic/*
• Google → google/*
• OpenAI → openai/*

Catalog presence is not a health check. A listed backend can still be warming, temporarily unavailable, or incompatible with a particular request option. The extension does not invoke every model at startup because probes are billable.

API compatibility

GPT-5.4 and newer GPT models use Lightning's /responses endpoint through pi's built-in OpenAI Responses implementation. The gateway rejects function tools combined with reasoning_effort on /chat/completions for these models. Switching APIs preserves reasoning rather than silently disabling it.

Older OpenAI models, Qwen, and other gateway models keep the Chat Completions path. The same routing policy applies to both discovered and offline models. Discovery also preserves the fallback models' extra thinking levels.

On upgrade, cached lai model entries are migrated to the current API and compatibility policy before Pi restores them, including during offline startup. Pricing, limits, and catalog freshness timestamps are preserved. No cache deletion or re-login is needed; credentials and other providers' caches are not modified.

Pricing and usage

Lightning AI returns OpenAI-style prompt, completion, total, and reasoning token usage. It does not return a dollar-cost field. The extension converts the live catalog's per-token input/output prices to pi's per-million-token rates, and pi calculates the cost locally.

Lightning AI currently publishes no cache-token prices for this endpoint, so cache read/write rates are zero.

Development

Try the extension without installing it:

pi -e ./extensions/lightning-ai.ts

Install development dependencies and run offline checks (no API calls):

npm ci --ignore-scripts
npm run check
npm test

Offline tests exercise pi's actual provider dispatch and request serializers, including tool declarations, tool-result replay, reasoning settings, catalog filtering, metadata, and fallback/discovery consistency. A package smoke test also packs the actual npm tarball and loads it through Pi's extension loader from an isolated temporary directory with no local dependencies or network.

Run the billable, opt-in live streaming tests:

LIGHTNING_API_KEY='...' npm run test:live

The default matrix is Qwen3.8, GPT-5.2, GPT-5.4 nano, GPT-5.5, GPT-5.6 Sol, and GPT-6.1 Sol. Each model must call a small tool, accept its result, report that result in its final answer, and return usage/cost data. Requests use low reasoning, at most 1,024 output tokens per turn, and a 60-second timeout. Live tests read only LIGHTNING_API_KEY, not pi's credential store or .env.

Override the matrix to verify other official catalog models:

LAI_TEST_MODELS='lightning-ai/gemma-4-31B-it,google/gemini-2.5-flash' npm run test:live

Missing catalog models fail explicitly; fallback availability is not proof that a backend is still offered. Live tests verify low reasoning only, not all thinking levels or multimodal capabilities.

Check exactly what npm would publish:

npm pack --dry-run

Release

PR checks run the offline suite, type-checking, dependency audit, and package check on Node.js 22 and 24. Development checks use Pi 1.1.0; packed-extension loading has also been verified with the minimum supported Pi 0.99.2. Live tests are never run automatically in CI.

After review and merging to main, publish from a clean checkout:

npm ci --ignore-scripts
npm run check
npm test
npm audit --audit-level=moderate
npm run check:pack
npm publish --access public

npm publish also runs type-checking and offline tests via prepublishOnly. The release version must be unused on npm; this branch prepares 0.1.1. Verify the published version with npm view pi-lightning-ai-provider version. Publishing requires npm package-owner credentials (and possibly an OTP).

License

MIT