pi-muse-spark

Pi extension for Meta Model API — adds Muse Spark 1.2, 1.2-contributor & 1.1 to pi coding agent via API key.

Fork note & credit: This is a maintained fork of seemethere/pi-meta-ai which provides the original implementation for most of the work. Original credit goes to seemethere/pi-meta-ai and its contributors. This fork fixes the false not authenticated warning on pi v0.84+ (where authStorage was removed) and keeps Muse Spark 1.2 models up to date. Upstream is currently unmaintained.

Meta Model API is OpenAI-compatible with Responses API at https://api.meta.ai/v1. This extension registers provider meta-ai using openai-responses for full agentic support: tool calling, parallel tools, streaming, reasoning effort, structured output, image input, prompt caching, 1M context.

Prerequisites

• Node.js >=18
• Pi coding agent >=0.84:

npm install -g --ignore-scripts @earendil-works/pi-coding-agent
pi --version # should be >=0.84

Install

As pi package (recommended):

pi install git:github.com/EclipseAditya/pi-muse-spark

# pin to a version:
pi install git:github.com/EclipseAditya/pi-muse-spark@v0.2.0

From source (dev):

git clone https://github.com/EclipseAditya/pi-muse-spark
cd pi-muse-spark
pi -e ./extensions/meta-model-api

Migrate from pi-meta-ai:

pi remove git:github.com/seemethere/pi-meta-ai
pi remove git:github.com/RooseveltAdvisors/pi-meta-ai
pi install git:github.com/EclipseAditya/pi-muse-spark

Quick start

1. Get API key: https://dev.meta.ai → API keys → Create. Key format LLM|...

2. Authenticate (pick one):

   Option A — Inside pi (recommended):

   /login → API key → Meta Model API → paste LLM|... key
   

   Option B — Env var (before launching pi):

   export MODEL_API_KEY="LLM|..."
   # Also supports META_API_KEY as fallback
   
   pi
   # or from source:
   MODEL_API_KEY="LLM|..." pi -e ./extensions/meta-model-api
   

   Env vars must be set before launching pi. If you set them after pi is running, run /reload or restart pi from the shell where var is exported.

3. Select model:

   /model → meta-ai/muse-spark-1.2-contributor
   # or meta-ai/muse-spark-1.2
   # or meta-ai/muse-spark-1.1
   

4. Verify:

   /meta status
   

5. Chat normally. Pi tools (read, bash, edit, write) work out of the box.

Commands

• /meta status — show provider, masked key status, auth source (stored/environment/runtime), active model
• /meta help — usage help
• /login — manage API keys
• /model — switch models

Features

• API: openai-responses
• Tool calling: parallel tools
• Reasoning: via pi thinking levels minimal, low, medium, high, xhigh → mapped to Meta minimal/low/medium/high (xhigh clamps to high)
• Structured output: JSON schema
• Input: text, image (PNG, JPEG, WebP, GIF)
• Caching: prompt caching automatic
• Context: 1M tokens, 64K max output
• Models:
  • muse-spark-1.2 — $1.25/$4.25 per M (balanced)
  • muse-spark-1.2-contributor — $0.10/$0.20 per M (cheap, community)
  • muse-spark-1.1 — free preview

Why this fork?

• Fixed bug: pi v0.84+ removed ctx.modelRegistry.authStorage. Original extension used authStorage.get() which always returned undefined, causing a false Meta Model API not authenticated. Run /login... warning even when ~/.pi/agent/auth.json was correctly configured. This fork uses getProviderAuthStatus() — the canonical pi API — so /login via auth.json is correctly detected.
• Active model detection: Now correctly detects muse-spark-1.2* as active, not just 1.1.
• Upstream credit: All provider/model logic, docs, and structure remain from seemethere/pi-meta-ai. See LICENSE (MIT).

Config reference (without extension)

If you prefer static config, create ~/.pi/agent/models.json — see models.json.example:

{
  "providers": {
    "meta-ai": {
      "baseUrl": "https://api.meta.ai/v1",
      "apiKey": "$MODEL_API_KEY",
      "api": "openai-responses",
      "models": [
        {
          "id": "muse-spark-1.2-contributor",
          "name": "Muse Spark 1.2 Contributor",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 1048576,
          "maxTokens": 64000,
          "cost": { "input": 0.1, "output": 0.2, "cacheRead": 0.002, "cacheWrite": 0.002 }
        }
      ]
    }
  }
}

Extension install is recommended for better login UX and future updates.

Troubleshooting

Model not visible in /model:

• Ensure extension is loaded: pi list should show git:github.com/EclipseAditya/pi-muse-spark
• Run /reload then /model
• Check /meta status shows Provider registered: yes

Meta Model API not authenticated (fixed in this fork):

• If you saw this with seemethere/pi-meta-ai on pi 0.84+, migrate to this fork (see Install).
• Run /login → API key → Meta Model API and paste LLM|... key
• Or export MODEL_API_KEY=LLM|... before launching pi, then /reload
• Verify with /meta status — should show Resolved: yes ✓ ready (source: stored)

Invalid key error:

• Key must start with LLM| from https://dev.meta.ai
• Re-create key at dev.meta.ai → API keys → Create
• Try /logout meta-ai then /login again

Env var not picked up:

• Export before launch: export MODEL_API_KEY=... && pi
• Inside running pi, /reload after export won't pick up new env if parent shell didn't have it — restart pi from shell where var is exported

Security

• Never commit real API keys, .env, or ~/.pi/agent/auth.json
• Use /login (stored in OS-protected auth.json) or env vars for keys
• Keys are displayed masked (e.g., LLM|...abcd) in /meta status
• See SECURITY.md

Development

npm install
npm run typecheck
pi -e ./extensions/meta-model-api/index.ts
# inside pi:
# /meta status
# /model → meta-ai/muse-spark-1.2-contributor

Structure:

pi-muse-spark/
  extensions/
    meta-model-api/
      index.ts      extension entry
  models.json.example
  package.json

Credits

• Original project: seemethere/pi-meta-ai — most code, docs, and model definitions originate there. Thank you!
• This fork: Bug fixes for pi 0.84+, Muse Spark 1.2 support, ongoing maintenance by EclipseAditya.
• License: MIT — see LICENSE (same as upstream).

License

MIT — see LICENSE