# Official AnyRouter provider for the pi coding agent

> The official pi extension registers AnyRouter as a provider, refreshes the live catalog and sets the compatibility flags pi would otherwise guess. One sk-ar-1- key, every chat model in the catalog.

- Date: November 22, 2026
- Category: guides
- HTML: https://anyrouter.dev/blog/pi-provider-anyrouter

## What the extension does

Pi reads custom OpenAI-compatible providers from `~/.pi/agent/models.json`. Hand-editing that file works, but the model list goes stale the day a model is added or retired.

The official extension registers `anyrouter` as a provider and refreshes the catalog from `https://anyrouter.dev/api/v1`, so `/model` in pi lists what AnyRouter can route that day. It delegates streaming to pi's built-in `openai-completions` implementation, so tool calls, usage accounting and cancellation keep working.

## Install it

Two options. The extension gives you the live catalog, the AnyRouter CLI gives you the shortest path:

```bash
pi install npm:@anyr/pi

# or from git
pi install git:github.com/AnyRouterHQ/pi-anyrouter

# or, with the AnyRouter CLI
curl -fsSL https://anyr.dev/cli.sh | bash
anyr pi
```

_Extension from npm or git, or the AnyRouter CLI._

The extension is published to npm under the `@anyr` scope as `@anyr/pi`, so `pi install npm:@anyr/pi` is the short form. The CLI writes the provider, sets `ANYROUTER_API_KEY` and starts `pi --provider anyrouter`, which is the fastest path when you do not want to manage the file yourself.

The source, the tests and the protocol limits are in [AnyRouterHQ/pi-anyrouter](https://github.com/AnyRouterHQ/pi-anyrouter).

## Why the compatibility flags matter

Pi decides how to talk to a provider from its host. For an unknown host it assumes a stock OpenAI endpoint and sends `store`, a `developer` role on reasoning models, `max_completion_tokens` and a long prompt-cache retention.

AnyRouter's chat endpoint does not document those fields. The extension sets pi's compat flags from each catalog row instead, so a request only sends parameters the model actually lists. Rows that list `reasoning_effort` get it, rows that list `reasoning` get the reasoning object, and rows that list neither get no reasoning field at all.

## What it does not do

Chat completions is the protocol this provider speaks. Claude Code's own request shape is Anthropic Messages, which pi's `anthropic-messages` client can emit, but that client also sends Claude-only fields that are not verified for the rest of the catalog, so no second provider is registered. Codex's shape is the Responses API, and pi's Codex client targets the ChatGPT backend, not this gateway.

Both still work through chat completions with their model ids, which AnyRouter accepts.

## Good to know

The catalog fetch is cached for five minutes per API key, matching the response `max-age`, and an offline refresh pass does not hit the network. Free models are `$0` per token on Go and up with no daily cap, and bring-your-own-key routes carry `$0` markup.

Pi has no MCP support, so there is nothing to auto-connect.
