pi-freerouter

Pi coding agent extension — auto-routes to free OpenRouter models

Packages

Package details

extension

Install pi-freerouter from npm and Pi will load the resources declared by the package manifest.

npm report
$ pi install npm:pi-freerouter
Package
pi-freerouter
Version
0.1.5
Published
Jun 7, 2026
Downloads
not available
Author
parkjangwon
License
MIT
Types
extension
Size
157.5 KB
Dependencies
0 dependencies · 1 peer
Pi manifest JSON
{
  "extensions": [
    "./index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-freerouter

Pi coding agent extension that routes every request through OpenRouter's free model tier — no paid API key needed beyond your OpenRouter account.

Quick start

1. Install

pi install npm:pi-freerouter

2. Set your OpenRouter API key

export OPENROUTER_API_KEY=sk-or-...

Free key at openrouter.ai/keys.

3. Start Pi

FreeRouter appears in the model picker and is set as the default automatically.

How free model routing works

OpenRouter exposes dozens of free models (models with a :free suffix). Each has its own rate limit — typically a few requests per minute per model. The trick is to spread load across all of them automatically.

Parallel racing

Every time Pi sends a request, pi-freerouter doesn't pick one model and hope for the best. It picks the next 3 available models from its sorted list and fires all three requests simultaneously.

Request arrives
    │
    ├── model A ──────────────────── first token ──▶ WINNER → stream to Pi
    ├── model B ───── (slower)                      → aborted
    └── model C ─── (even slower)                   → aborted

Whichever model emits its first token wins. The other two are immediately cancelled. Pi sees a single clean stream — it has no idea a race happened.

Automatic fallback

When a model hits its rate limit (HTTP 429) or fails (5xx), it's marked as exhausted and skipped for 90 seconds. The next batch of 3 models takes over. Timed-out models (no response in 30 seconds) recover faster — after 15 seconds.

Batch 1: [model A, model B, model C]  → all hit quota
Batch 2: [model D, model E, model F]  → model D wins
         ↑ model A–C recover after 90s and rejoin the pool

Provider priority

Free models are sorted so the lowest-latency inference providers are always tried first:

  1. Groq
  2. Cerebras
  3. Fireworks
  4. Together
  5. Mistral
  6. Everything else (sorted by context window ascending)

Model list refresh

The list of available free models is fetched at startup and refreshed every hour in the background, so long-running Pi sessions automatically pick up newly added models.

Requirements

License

MIT