@lowrouter/pi-provider0.1.4

LowRouter provider for Pi

Adds LowRouter as a model provider in Pi.

LowRouter is an OpenAI-compatible router run from the EU. An auto/<creator>/<model> ID can be served by several European or global hosts. LowRouter prefers low-carbon regions and fails over between them. Maintained by the LowRouter team.

Install

Requires Pi 1.3.0 or newer and Node.js 22.19 or newer.

pi install npm:@lowrouter/pi-provider

LowRouter's models show up in Pi without a key. To send requests, get a key at lowrouter.ai/app/keys, then either export it or store it with /login inside Pi (pick LowRouter):

export LOWROUTER_API_KEY=sk-lr-...

Pick a model:

pi --list-models lowrouter
pi --model lowrouter/auto/mistralai/mistral-large-2512

or /model in a session.

Models

The provider reads GET /v1/models?currency=usd&function_calling=yes&modality=chat when Pi starts, when the model picker refreshes and on pi update --models. Without a key it reads the public list. With a key the key's catalog allowlist applies and the account's alias/ models are added. Pi keeps the last list for offline starts. The picker lists:

• auto/<creator>/<model> IDs and your account's alias/<name> models;
• chat models with tool calling only, since Pi acts through tools.

Pinned routes (<host>/<creator>/<model>) and jurisdiction-scoped IDs (auto/<creator>/<model>/eu) stay out of the list. To use one, add it in models.json as described in the LowRouter guide.

Claude models in the picker (auto/anthropic/<model>) go through Pi's Anthropic API on LowRouter's /v1/messages, so cache_control breakpoints and tool calls keep Anthropic's own format. Every other model uses /v1/chat/completions.

Context windows, output limits and input types come from /v1/models. A model reasons when its default routes publish reasoning_effort levels there; Pi's thinking levels then map onto exactly those levels, and a level the routes don't take is not offered. Other models get no effort. A Claude model whose reasoning_mode is adaptive takes the level as an effort; one in budget mode gets a thinking-token budget for it. A model without a published context window gets 32,768 tokens, so Pi compacts early rather than sending a prompt the model rejects. A model without an output limit gets 8,192.

Until a live or cached list is in, Pi uses the snapshot shipped with the release, and it stays when Pi is offline (--offline or PI_OFFLINE) with nothing cached. The snapshot is generated from the same endpoint by npm run gen:catalog; nothing in it is written by hand.

Without a key

Pi lists only providers it counts as configured, so without a key the provider still reports LowRouter as configured, with no credential. The models show in /model and pi --list-models, but a request fails before it is sent, with a message to run /login or set LOWROUTER_API_KEY. Two things follow from Pi counting LowRouter as configured:

• If no default model is saved and no configured provider has a Pi default model (a custom models.json provider has none), a fresh Pi can pick a LowRouter model as its default, and the first prompt shows that message. With no other provider at all, it replaces Pi's "No models available" notice.
• A session, or Pi's saved default model, on a LowRouter model restores without a key and fails at the first prompt, instead of switching to another provider.

Cost

LowRouter bills in euros, for the route that served the request. Pi tracks cost in US dollars, so the provider reads USD prices (?currency=usd). An auto/ model has several routes at different prices, and the provider shows the dearest one, so Pi's estimate is an upper bound for input and output tokens. Cache tokens are different: /v1/models publishes no cache-write price and a cache-read price for only some models, and Pi counts an unpublished price as zero. Claude caches the prompt prefix on every turn, so its estimate can come out below the bill. What you were charged is in each response's usage.cost (EUR) and on lowrouter.ai/app.

Development

npm install
npm run typecheck
npm test
npm pack --dry-run

Try a local checkout without installing it: pi -e ./dist/index.js after npm run build.

Live tests need a key and explicit opt-in. They run discovery, one streamed completion and one tool call on each of six models, a thinking-level turn on three, and a two-turn prompt-cache check on Claude Haiku 4.5:

LOWROUTER_LIVE_TEST=1 LOWROUTER_API_KEY=sk-lr-... npm run test:live

To release:

1. Bump version in package.json and refresh the snapshot: npm run gen:catalog. It reads the public catalog, without a key.
2. Push to main, then run the Publish workflow from the Actions tab, first with dry_run on. It stages the version on npm with provenance through trusted publishing, with no stored token.
3. Approve the staged version with 2FA on npmjs.com, or npm stage approve <stage-id> (npm stage list @lowrouter/pi-provider shows the id). Until then it is not installable.

License

Apache-2.0