№ 0102GitHub
Muse Spark Anywhere
A set of small fixes that let Muse Spark run in pi, Agent Orchestrator, opencode and other OpenAI-compatible tools, handling its non-standard SSE event and missing model-catalog entries.
# Muse Spark Anywhere Use **Meta's Muse Spark** (the model behind the `muse` CLI / Meta Model API) in **pi**, **Agent Orchestrator**, **opencode**, and any OpenAI-compatible tool — with full reasoning-effort control and no secrets in your configs. Muse Spark speaks an OpenAI-compatible API (`https://api.meta.ai/v1`), but two things trip up most tools: 1. **Streaming quirk** — it injects a non-standard SSE event (`event: response.subscription_usage`) that strict OpenAI parsers reject. 2. **Discovery gaps** — it isn't in [models.dev](https://models.dev), so tools that generate their catalog from there (pi, opencode, Kilo Code) don't list it, and some UIs (Agent Orchestrator) hardcode an older model list. This repo collects small, focused fixes for each. ## What's here | Path | What it does | |------|--------------| | [`pi/`](pi/) | Drop-in `~/.pi/agent/models.json` + a Keychain key reader. Muse Spark in **pi today**, with `minimal…max` thinking levels. pi handles the streaming quirk natively. | | [`proxy/`](proxy/) | A ~120-line, dependency-free Node proxy that strips the non-standard SSE event, so **opencode / any OpenAI GUI** works. Point them at `http://127.0.0.1:8788/v1`. | | [`agent-orchestrator/`](agent-orchestrator/) | A patch adding `muse-spark-1.3` (+ contributor) to Agent Orchestrator's muse model catalog. | | [`pi-provider/`](pi-provider/) | Draft for an upstream pi PR: a built-in `meta` provider (mirrors the existing openai-compatible providers). | ## The clean, ecosystem-wide fix Because pi, opencode, and Kilo Code all build their catalogs from **models.dev**, the highest-leverage contribution is adding a **Meta provider + Muse Spark models** (with reasoning options) to models.dev. Then every one of those tools gets Muse Spark natively. The `pi-provider/` draft is the pi-side wiring that pairs with it. ## Auth & safety No key material is stored in any file here. The pi config and the proxy read the key from the macOS Keychain (service `ai.meta.dev.credentials`, written by the `muse` CLI) at request time. If you don't use the `muse` CLI, set `META_API_KEY` in your environment instead. This is developer tooling for your **own** Meta Model API access — mind Meta's terms for how your plan/key may be used. ## Quick start (pi) ```bash cp pi/models.json ~/.pi/agent/models.json cp pi/meta-key.sh ~/.pi/agent/meta-key.sh && chmod +x ~/.



ChatForm
Tgmlabs