pi-uberllm0.2.1

pi-uberllm

A Pi extension that adds UberLLM as a model provider.

It reads UberLLM's live model catalog and registers it with Pi, so every model UberLLM serves shows up in the /model picker. There is no model list to maintain by hand.

Provider	UberLLM surface	Models
uberllm	OpenAI-compatible — https://api.uberllm.dev/v1	every chat model UberLLM serves
uberllm-anthropic	Anthropic-compatible — https://api.uberllm.dev	anthropic/* (Claude)

The catalog is refreshed at startup and on every new session, so models that appear or disappear on UberLLM's marketplace show up without restarting Pi.

Install

pi install git:github.com/Milor123/pi-uberllm

Pin a version instead of tracking the default branch:

pi install git:github.com/Milor123/pi-uberllm@v0.2.1

To try it for one invocation without adding it to your settings:

pi -e git:github.com/Milor123/pi-uberllm

Setup

Set your UberLLM API key once. It starts with ull_ and is created at https://uberllm.dev/dashboard/keys.

# Windows (persistent, user scope)
setx UBERLLM_API_KEY "ull_your_key_here"

# macOS / Linux
export UBERLLM_API_KEY="ull_your_key_here"

Restart your terminal and Pi afterwards. Then run /model and pick anything under UberLLM.

Is it plug and play? Yes, after those two steps: install the package and set the key. You do not need to edit models.json, settings.json or auth.json, and the key is never written to a config file — the extension declares apiKey: "$UBERLLM_API_KEY" and Pi resolves it from the environment.

Alternatively, run /login and pick uberllm to store the key through Pi instead of using an environment variable.

Why two providers

UberLLM speaks both wire formats, and the split is not cosmetic — it decides whether prompt caching works.

Anthropic-style prompt caching needs cache_control markers on the request. Pi emits those itself on the anthropic-messages API and does not on the OpenAI-compatible one. Routing Claude through /v1/messages keeps caching in Pi's hands instead of relying on the aggregator to inject it, which is the difference between paying cache-read rates and paying full input rates.

UberLLM publishes cache_read_usd and cache_write_usd per model, and both are mapped into Pi's model cost. Without that mapping Pi reports cached tokens at the input rate and the session total does not match the UberLLM bill.

Prompt cache warming is deliberately off. A warm-up request is a billed request against prepaid credit, so this extension leaves promptCache unset. To opt in, add promptCache: { short: 300, long: 3600 } to the model in src/catalog.ts.

To see whether caching is actually happening, set "showCacheMissNotices": true in Pi's settings.json.

Compatibility notes

Pi's guidance is to set compatibility flags only for differences verified against the real service, never because an endpoint advertises compatibility. This extension sets exactly one, in src/catalog.ts:

• supportsDeveloperRole: false on the OpenAI-compatible surface. UberLLM rejects OpenAI's developer role with developer is not one of ['system', 'assistant', 'user', 'tool', 'function'], and Pi auto-detects the flag as enabled for a base URL it does not know.

Reasoning uses the OpenRouter wire format, because UberLLM documents its catalog as OpenRouter-shaped and uses OpenRouter's names in supported_parameters.

Every Pi thinking level was verified against the live service:

Pi level	Sent as	Result
off	reasoning: { effort: "none" }	accepted, zero reasoning tokens
minimal … max	reasoning: { effort: "<level>" }	accepted

No thinkingLevelMap is needed, because every Pi level maps 1:1. When a model does not offer the level you picked, UberLLM returns the nearest level it does offer instead of failing. Note that UberLLM rejects the spellings "off" and "disabled" with HTTP 400 — Pi sends "none", so this never triggers, and the "openai" format has no equivalent for a disabled level anyway.

Models are filtered to type: "chat", available: true and nsfw: false. UberLLM's available flag is live marketplace state: a model with no seller right now answers 503 no_available_provider, so it is not offered.

Development

npm test

The mapping logic lives in src/catalog.ts with no network access and no Pi runtime dependency, so it is unit tested directly. src/index.ts owns fetching and provider registration.

To work on it against your own Pi install, point a directory at the repository. On Windows a junction works and needs no privileges:

mklink /J "%USERPROFILE%\.pi\agent\extensions\uberllm" "C:\path\to\pi-uberllm"

On macOS and Linux:

ln -s /path/to/pi-uberllm ~/.pi/agent/extensions/uberllm

A single-file symlink does not work for this extension: src/index.ts imports ./catalog, which must resolve inside the loaded directory.

License

MIT — see LICENSE.