pi-meta-provider

Meta Model API provider extension for pi. Adds Meta's Muse Spark models — Meta's family for agentic and coding workflows — to pi through the OpenAI-compatible Chat Completions endpoint.

Models

Model	Tier	Context	Price (in/out/cached, per 1M tokens)
muse-spark-1.1	Standard	1,048,576	$1.25 / $4.25 / $0.15
muse-spark-1.2	Standard	1,048,576	$1.25 / $4.25 / $0.15
muse-spark-1.2-contributor	Contributor	1,048,576	$0.10 / $0.20 / $0.002

All Muse Spark models support up to 131,072 output tokens and always reason — reasoning_effort maps to pi's thinking levels (minimal → xhigh; "off" omits the parameter since "none" is rejected by the API).

Install

Set your API key (create one in the Model API dashboard):

export MODEL_API_KEY="LLM|..."

As an extension (auto-discovered)

mkdir -p ~/.pi/agent/extensions
git clone https://github.com/franktheglock/pi-meta-provider ~/.pi/agent/extensions/pi-meta-provider

As a pi package

pi install git:github.com/franktheglock/pi-meta-provider

Usage

pi --model meta/muse-spark-1.2

Select it any time with /model, and dial reasoning with the thinking-level controls (e.g. /thinking high).

How it works

The extension registers the meta provider via pi.registerProvider():

• API: openai-completions — pi streams over POST /v1/chat/completions (SSE), the protocol Meta documents as drop-in compatible with OpenAI clients and agent CLIs.
• Auth: Authorization: Bearer from MODEL_API_KEY.
• Reasoning: reasoning_effort sent at the top level, mapped from pi's thinking levels via thinkingLevelMap.
• Compatibility flags: supportsDeveloperRole (Meta's developer instruction role), maxTokensField: "max_completion_tokens", supportsLongCacheRetention (sends prompt_cache_retention: "24h" when long retention is enabled), and supportsReasoningEffort.
• Costs: per-tier pricing with prompt_tokens_details.cached_tokens reported as cache reads.

Resources

• Meta Model API overview
• Chat Completions protocol
• [Pricing and rate limits](https://dev.meta.ai/docs/pr