Use Pi with abliteration.ai

Use the Pi coding agent with abliteration.ai through a custom OpenAI-compatible provider, with session affinity enabled for prompt-cache routing.

Pi is a terminal coding agent. Configure abliteration.ai as a custom OpenAI-compatible provider and enable Pi’s session-affinity headers so requests from the same conversation are routed together for better prompt-cache reuse.

Install

Pi requires Node.js 22.19 or later.

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

Set your API key

Get an ak_... key from the console, then export it in your shell:

export ABLIT_KEY="ak_YOUR_API_KEY"

Configure the provider

Add the following provider to ~/.pi/agent/models.json:

{
  "providers": {
    "abliteration": {
      "baseUrl": "https://api.abliteration.ai/v1",
      "apiKey": "$ABLIT_KEY",
      "api": "openai-completions",
      "compat": {
        "sendSessionAffinityHeaders": true
      },
      "models": [
        {
          "id": "abliterated-model-large-v2",
          "name": "Abliterated Model Large V2",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 1000000,
          "maxTokens": 999990,
          "cost": {
            "input": 5,
            "output": 5,
            "cacheRead": 0.5,
            "cacheWrite": 5
          }
        },
        {
          "id": "abliterated-model",
          "name": "Abliterated Model",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 262144,
          "maxTokens": 262134,
          "cost": {
            "input": 3,
            "output": 3,
            "cacheRead": 0.3,
            "cacheWrite": 3
          }
        }
      ]
    }
  }
}

Keep sendSessionAffinityHeaders enabled. Pi auto-detects this behavior only for known providers, so the custom abliteration.ai provider must enable it explicitly. Pi then sends its session ID in x-session-affinity, which abliteration.ai uses as a cache-routing hint. See Pi’s provider compatibility reference and prompt caching on abliteration.ai.

Do not replace this setting with one static custom header. Pi derives a stable value for each session, keeping unrelated conversations in separate affinity groups.

Start Pi

Run pi, enter /model, and select the abliteration provider and your preferred model. Use Abliterated Model when you need image input; Abliterated Model Large V2 is text-only.