pi-provider-lemonade

Pi extension that turns your local Lemonade Server into a model provider. It discovers the models you downloaded on the server automatically — add a model there, and it shows up in pi's /model list. No models.json editing.

What you need

• pi installed (verified with pi 1.3.0; on a newer pi release, streaming should still work — and if it misbehaves, see Troubleshooting for a one-line switch).
• Lemonade Server installed and running, with at least one model downloaded:
  • Web UI / Desktop App: tick "Downloaded Only" to see what's ready.
  • CLI: lemonade list --downloaded.

Quick start

1. Install the extension:

   pi install npm:pi-provider-lemonade
   

2. Start pi (pi). If your Lemonade Server runs on this machine with default settings, you're done — the extension finds it on its own.

3. In the TUI, run /model (or press ctrl-l) and pick any lemonade/… entry. Chat away.

Everyday commands

Notes:

• The /model dialog refreshes the catalog in the background while it is open, so a model you just downloaded usually appears the next time you open the dialog — no /reload needed.
• Images: models that report the vision capability accept images pasted or dragged into the prompt. (Vision models without an mmproj companion file fail on image input — a known Lemonade limitation, lemonade-sdk/lemonade#955.)
• Thinking: models with a thinking toggle (e.g. Qwen3, gpt-oss) show pi's thinking-level selector. This is configured automatically per model family; if a specific model's toggle misbehaves, you can override it in models.json (see Fine-tuning models).

Connecting to the server

By default the extension talks to http://127.0.0.1:13305 — the Lemonade Server's default address. If your server runs locally with defaults, skip this section entirely.

Remote or non-default server — two ways:

1. /login lemonade in the TUI (preferred). It shows a list of known servers — the one from a previous login, the address from the environment variables (below), and any servers found by a network scan — each with a live online/offline status. Pick one (or use Scan local network…, or Enter custom URL…), then enter the API key if your server asks for one. Both are stored in pi's credential store.

2. Environment variables. Set these before starting pi:

   Variable	Default	Purpose
   LEMONADE_HOST	127.0.0.1	Server host — bare host (192.168.1.100), host:port, or a full http:///https:// URL
   LEMONADE_PORT	13305	Server port (used unless LEMONADE_HOST already includes one)
   LEMONADE_API_KEY	—	API key sent to the server
   LEMONADE_ADMIN_API_KEY	—	Admin key; wins over LEMONADE_API_KEY when both are set, exactly like the lemonade CLI

A /login credential takes priority over the environment variables.

Security note: unless the server address is an https:// URL, requests — including the API key — travel unencrypted. That's fine on your own machine, but for a server on another host use an https:// URL (or a tunnel), and prefer /login lemonade over environment variables so the key isn't readable by every process in your shell.

Troubleshooting

Fine-tuning models

Everything numeric or model-specific can be overridden per model in ~/.pi/agent/models.json, just like for any other pi provider — no restart needed, open /model afterwards:

{
  "$schema": "https://pi.dev/schemas/models.schema.json",
  "providers": {
    "lemonade": {
      "modelOverrides": {
        "Qwen3-4B-GGUF": {
          "name": "Qwen3 4B (High Ctx)",
          "contextWindow": 131072,
          "maxTokens": 32768
        }
      }
    }
  }
}

Useful override fields: contextWindow, maxTokens, reasoning (force the thinking selector on/off), compat.thinkingFormat (how thinking controls are sent to the server — change it only if a model's toggle does nothing or the server rejects requests), input, and cost.

The extension discovers sensible defaults for all of these from the server: contextWindow from the server's model info (fallback 128,000), maxTokens 16,384 capped to what fits in the context window, cost from the server for cloud models (0 for local ones), and input: ["text", "image"] for vision models.

Extension settings (optional)

All of these have working defaults — most users never need this file. It lives at ~/.pi/agent/pi-provider-lemonade.json (or the directory in PI_CODING_AGENT_DIR):

{
  "discoveryTimeoutMs": 20000,
  "streamingMode": "extension",
  "startupScan": true
}

How model discovery works

The extension lists the server's models (GET /v1/models), asks each one for its capabilities (POST /api/show), and registers only chat models that support both completion and tools (embedding, image-generation, and transcription models are skipped). Results are sorted: pinned models first, then loaded ones, then other downloaded local models, then unloaded cloud models. Only models with downloaded: true are included — download models with the Lemonade app or CLI, not through pi.

The extension covers Lemonade's OpenAI-compatible chat surface only. Downloading models, image generation, and other server features are done with the Lemonade app or CLI.

Feedback and license

Bugs and feature requests: open an issue (pull requests are not accepted). License: MIT.