@aistupidlevel/pi-provider0.1.0

AI Stupid Level for Pi

Stop choosing AI models manually. Let live benchmarks choose the best model for coding, reasoning, or agent tasks—using your own provider keys, with zero token markup.

A Pi provider for the AI Stupid Level Smart Router. Instead of a fixed model you pick a routing mode such as auto-coding, and the router sends each request to the model with the best measured score on AI Stupid Level's benchmarks, among the providers you connected (OpenAI, Anthropic, Google, DeepSeek, Kimi, GLM). It calls that model with your own provider key.

Install

pi install npm:@aistupidlevel/pi-provider

Set up

1. Create a router key at https://aistupidlevel.info/router/keys (it starts with aism_).

2. On the same site, add the provider keys the router may use.

3. Give Pi the router key. In Pi, run /login, choose Sign in with an API key, then AI Stupid Level, and pick a mode with /model afterwards. Or set an environment variable:

   export AISTUPIDLEVEL_API_KEY=aism_...   # ASL_ROUTER_KEY works too
   

   A key saved with /login is stored in ~/.pi/agent/auth.json and takes precedence over the environment.

Use

pi --model aistupidlevel/auto-coding

Or open /model and pick an AI Stupid Level mode; Ctrl+S there makes it the default for new sessions. The footer shows which model answered the latest request, for example ASL → claude-sonnet-5-5, and Pi records it with each answer in the session.

Routing modes

Model	Sends each request to
auto-coding	the best score on the coding benchmark (the default)
auto-tooling	the best score on the tool-use benchmark; suits agent work
auto-reasoning	the best score on the reasoning benchmark
auto-best	the best combined score
auto-task	the benchmark that matches each request: tools, reasoning, code or overall
auto-value, auto-value-coding, auto-value-reasoning, auto-value-tooling	the most benchmark points per measured dollar, among models within 5 points of the best
auto-fastest-quality	the fastest model within 5 points of the best
auto-consistent	the model whose score varies least, within 5 points of the best
auto-cheapest, auto-fastest	the lowest price per token, or the lowest latency, with no quality bar
auto	the strategy saved in your router preferences
auto-split	your own traffic split, set in your router preferences

When the leaders are statistically tied, the router picks the one that cost least per task in its measurements; you can choose the fastest or the highest score instead in your router preferences. A conversation stays with the model that answered it, so an agent loop is not handed to a different model mid-task. A model id such as aistupidlevel/claude-opus-5-5 pins one model.

The model list comes live from the router (GET /v1/models with your key): Pi refreshes it in the background when it starts interactively, or when you open /model, and keeps the last list it fetched when the router cannot be reached. Before the first refresh, the main routing modes are listed.

Full documentation: https://aistupidlevel.info/router/docs (routing modes: #strategies).

What is measured

AI Stupid Level runs its own coding, reasoning and tool-use benchmarks on the models the router chooses from and publishes the results at https://aistupidlevel.info.

• Each model ranks on the average of its last five measured runs, never on a single run.
• Value is measured, too: the benchmark score divided by what one identical benchmark run cost that model, not its list price.
• Drift-aware routing, on by default: a model whose drift detector has fired for the skill your mode relies on, or whose hourly canary is failing, moves behind every unflagged model.

Your keys, zero markup

The router calls the chosen model with the provider key you saved on AI Stupid Level, so the provider bills you directly, at its own prices. AI Stupid Level adds 0% markup to provider tokens on every plan, and the free account includes 1,000 router requests a month (pricing).

That is also why these models show a cost of $0 in Pi: the price depends on which model answered, and it appears on your provider's bill.

Text and tool calls only: the router does not route image input.

Configuration

Variable	Purpose
AISTUPIDLEVEL_API_KEY	Router key (aism_…), when none was saved with /login
ASL_ROUTER_KEY	Read when AISTUPIDLEVEL_API_KEY is not set

To reach the router through another endpoint, such as a company proxy, set its baseUrl in Pi's ~/.pi/agent/models.json:

{ "providers": { "aistupidlevel": { "baseUrl": "https://proxy.example.com/v1" } } }

Every model is declared with a 262,144-token context window and 64,000 output tokens: the smallest limits across the router's pool, so they hold for any model it picks.

What this extension does (disclosure)

• Network calls: it registers one model provider. When you use it, Pi sends your prompts, tool definitions, tool results and your router key to https://aistupidlevel.info/v1. The extension itself calls GET /v1/models with your key to list models. The router forwards each request to the provider it routes to, using your own key for that provider.
• Nothing else: no tools, commands, files read or written, background processes or telemetry. It reads the X-AISM-Model response header to show the answering model in the footer.
• Prompt logging: prompts are not stored by AI Stupid Level unless you enable prompt logging for a key in your account. See https://aistupidlevel.info/privacy.

Development

npm install
npm test                               # typecheck and unit tests
ASL_ROUTER_KEY=aism_... npm test       # adds live tests in sandbox mode: real routing, mock answers, nothing spent
pi -e ./index.ts --model aistupidlevel/auto-coding

Pi loads index.ts directly; there is no build step. Built and tested against Pi 1.4.0 (@earendil-works/pi-coding-agent) on Node 22.

Licence

MIT