pi-edgee

Pi provider extension for Edgee, the Agent Gateway that compresses, routes, and observes LLM requests.

Learn more about Edgee in the official documentation.

Installation

pi install npm:@aliou/pi-edgee

Set the API key before use:

export EDGEE_API_KEY=...

What it provides

• Registers an edgee provider in Pi.
• Uses Edgee’s OpenAI-compatible Gateway API directly at https://edgee.io/v1, not an SDK.
• Fetches models from GET https://edgee.io/v1/models when EDGEE_API_KEY is available. Model ids use the {provider}/{model} format (e.g. openai/gpt-5.2); the owned_by field exposes the upstream provider.
• Falls back to a hardcoded model list (enriched with pricing/limits from models.dev) when model fetching is unavailable.
• Uses Pi’s native openai-completions provider path, so tool calls, streaming usage, and context-overflow normalization all work.
• Normalizes upstream context-overflow errors through Edgee into Pi’s context_length_exceeded signal.

Two APIs

Edgee exposes two separate APIs. This extension only needs the Gateway API:

• Gateway API (https://edgee.io) — LLM requests and model listing. Authenticated with an sk-edgee-... API key (EDGEE_API_KEY). Endpoints: /v1/models, /v1/chat/completions, /v1/messages, /v1/responses, /v1/compress, /v1/count_tokens.
• Console API (https://api.edgee.app) — organization management: usage/cost export, gateway API-key management, BYOK. Requires a separate ek_live_... Console token and an org id, not the gateway key. Not used by this extension.

Development

pnpm install
pnpm typecheck
pnpm lint
pnpm test

Layout

• extensions/provider/: Pi extension entry point.
• src/edgee/: Edgee model mapping and API helpers.

Notes

Edgee compresses input tokens before they reach the upstream provider, so actual billed cost is usually lower than the cost figures embedded in the model metadata (which reflect upstream list prices from models.dev). Per-request savings are reported in the compression field of Edgee’s responses and are not reflected here.