Docs · Guide

Pi on any model

Pi, the pi coding agent, takes providers of your own in its models.json. magpie adds itself there, so pi can run on any model you have in magpie: DeepSeek, GLM, Qwen, an OpenAI or Anthropic key, or a ChatGPT or Copilot subscription, with no key pasted into pi.

Pi keeps its settings in ~/.pi/agent, or in the folder $PI_CODING_AGENT_DIR names. magpie adds a provider named magpie to models.json, with the gateway’s address, a key of its own and the models you have in magpie, so they join pi’s model picker. Then it sets defaultProvider and defaultModel in settings.json to your pick. A Pi in WSL is set up the same way.

Connect Pi

1. Add providers in magpieIn the Providers tab, add an API key or sign in to a subscription.
2. Pick a model on Pi’s rowIn the Agents tab, choose a model under magpie.
3. Run piIt starts on that model, and every model of magpie’s is in its model picker.

magpie pi deepseek/deepseek-v4-pro    # Pi on DeepSeek
magpie pi                             # what Pi runs on now

What each model gets

• Its own APIEach model is asked on its provider’s native API: GPT and ChatGPT models on Responses, Claude on Messages, the rest on Chat Completions. A Claude model that thinks only adaptively gets forceAdaptiveThinking.
• Name, context and repliesThe name pi shows, reasoning for models that think, contextWindow, image input where the model takes images, and maxTokens. A maxTokens you typed in yourself stays.
• Thinking levelThe level you pick in magpie goes in defaultThinkingLevel, held to the levels Pi offers.

Pi sends its session with each request (sendSessionAffinityHeaders), so magpie’s usage shows each pi session and Record gateway conversations keeps them. A value you set there yourself stays.

The Ctrl+P list

Pi cycles models with Ctrl+P through enabledModels, the list /scoped-models edits. If you have such a list, magpie adds the model you picked to it. If you don’t, magpie never makes one, so Ctrl+P keeps cycling every model.

Decision and image models

magpie also puts an extension, extensions/magpie.ts, in Pi’s folder. It adds magpie’s decision models and image models, for codemode. It is there while the magpie provider is, and goes with it.

Your own files

Before its first write, magpie copies settings.json and models.json to .before-magpie files beside them. Your own providers and settings stay as they are, and Pi’s row lists your models beside magpie’s.

Switch back

Pick one of your own models on the row, take the row off magpie, or run magpie pi default. If you changed nothing since, magpie puts both files back byte for byte. Otherwise it removes only its own provider and keeps your changes.

FAQ

Can Pi run on DeepSeek or GLM?

Yes. Any model you have in magpie is in pi’s model picker: DeepSeek, GLM, Qwen, GPT, Claude and more.

Can Pi use my ChatGPT subscription?

Yes. Sign in to ChatGPT in magpie and pick a GPT model under magpie on Pi’s row. pi sends its requests to magpie, which uses the subscription on the Responses API.

Where does magpie write if I set PI_CODING_AGENT_DIR?

In that folder, with ~ expanded. A relative path is ignored, and magpie uses ~/.pi/agent.

Does it work with Pi in WSL?

Yes. magpie sets up a Pi in a WSL distro too, its provider pointing at the gateway as the distro reaches it.

More agents on the same models: Models in OpenCode, Models in Claude Code and Kimi Code on any model. Something fails? See Fixing common errors, or ask on Discord.