Pi LLM Gateway Extension

A Pi extension that adds LLM Gateway as a model provider, giving you access to 160+ chat models from OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, and many others through a single OpenAI-compatible API.

Installation

Get API Key

Sign up at llmgateway.io to get an API key.

Configure Credentials

Add your API key to ~/.pi/agent/auth.json (recommended):

{
  "llmgateway": { "type": "api_key", "key": "your-api-key-here" }
}

Or set an environment variable:

export LLMGATEWAY_API_KEY="your-api-key-here"

Install Extension

# From npm
pi install npm:@mcowger/pi-llmgateway

# From git
pi install git:github.com/mcowger/pi-llmgateway

# Local development
pi -e ./src/extensions/provider/index.ts

Usage

Select llmgateway as your provider and choose from available models:

/model llmgateway gpt-5
/model llmgateway claude-opus-4-8
/model llmgateway gemini-2.5-pro
/model llmgateway auto

The special auto model lets the gateway pick the best provider and model for each request based on the configured routing strategy.

Settings

Configure with /llmgateway:settings:

Setting	Values	Default	Description
routing	auto, price, throughput, latency	auto	Provider routing strategy. Note: coding (dev) plans only support auto and price.
webSearch	enabled, disabled	disabled	Enable native web search for models that support it.
baseUrl	URL string	https://api.llmgateway.io/v1	API base URL for self-hosted instances.
includeDeactivated	include, ignore	ignore	Show deactivated models (may stop working at any time).

Settings can also be changed with pi config.

Compat values

The extension sends all requests in standard OpenAI Chat Completions format. The LLM Gateway normalises everything server-side:

• max_tokens is used for all models (the gateway translates to max_completion_tokens for GPT-5/o-series internally)
• developer role is not used (system role is sent for all models)
• Reasoning is controlled via top-level reasoning_effort (none|minimal|low|medium|high|xhigh|max)
• reasoning_content on assistant messages is not required (the gateway reconstructs it for upstreams that need it)
• Prompt caching is managed by the gateway

Self-hosted instances

Change the base URL in settings to point to your own LLM Gateway deployment:

/llmgateway:settings → Connection → Base URL → https://your-gateway.example.com/v1

Restart Pi after changing the URL.

Updating the Model Snapshot

Models are seeded from a hardcoded snapshot (src/extensions/provider/models/static.ts) and refreshed from the live /v1/models API on each session start. To update the snapshot (e.g. after significant model changes):

1. Run bun run test — it fetches the live API and reports drift
2. Regenerate the snapshot by running the generator script
3. Re-run bun run test to confirm no drift

Development

git clone https://github.com/mcowger/pi-llmgateway.git
cd pi-llmgateway
bun install

Commands

bun run typecheck   # TypeScript type check
bun run lint        # Biome lint
bun run format      # Biome format (auto-fix)
bun run test        # Vitest (includes live API drift check)
bun run gen:schema  # Regenerate schema.json from config.ts

Requirements

• Pi coding agent v0.67.68+
• LLM Gateway API key (configured in ~/.pi/agent/auth.json or via LLMGATEWAY_API_KEY)

Links

• LLM Gateway
• LLM Gateway Docs
• LLM Gateway GitHub
• Pi Documentation