@zerodawn/pi-litellm-discovery

Pi extension that fetches models from a LiteLLM gateway and registers them as a custom provider at startup.

Install

Pi package install:

pi install npm:@zerodawn/pi-litellm-discovery

Optional helper command install via npm:

npm install -g @zerodawn/pi-litellm-discovery

That exposes:

pi-litellm-devenv

Local development from this monorepo:

pi -e ./packages/litellm-discovery

Repo-local helper script:

./packages/litellm-discovery/bin/pi-litellm-devenv

Required environment

• PI_LITELLM_DISCOVERY_BASE_URL: LiteLLM base URL, for example https://litellm.example.com/v1
• one of:
  • PI_LITELLM_DISCOVERY_API_KEY
  • PI_LITELLM_DISCOVERY_API_KEY_CMD

Optional:

• PI_LITELLM_DISCOVERY_PROVIDER: provider id shown in Pi, default litellm-discovery
• PI_LITELLM_DISCOVERY_NAME: provider label shown in Pi, default LiteLLM Discovery

Example

Direct key:

export PI_LITELLM_DISCOVERY_BASE_URL='https://litellm.lan.kappen.io/v1'
export PI_LITELLM_DISCOVERY_API_KEY='...'
pi -e ./packages/litellm-discovery

Shell command for the key:

export PI_LITELLM_DISCOVERY_BASE_URL='https://litellm.lan.kappen.io/v1'
export PI_LITELLM_DISCOVERY_PROVIDER='litellm-devenv'
export PI_LITELLM_DISCOVERY_NAME='LiteLLM Devenv'
export PI_LITELLM_DISCOVERY_API_KEY_CMD='infisical run --log-level error --silent --project-config-dir=/data/private/devenv --env=prod --path=/devenv/litellm -- sh -c '\''printf %s "$LITELLM_MASTER_KEY"'\'''
pi -e ./packages/litellm-discovery

Helper script with the same defaults:

./packages/litellm-discovery/bin/pi-litellm-devenv

Installed helper command after npm install -g:

pi-litellm-devenv

Override the devenv repo path if needed:

DEVENV_REPO=/some/other/devenv pi-litellm-devenv

Alias

If you want this as your normal shortcut, add this once to your shell profile:

alias pi-dev='pi-litellm-devenv'

Then use:

pi-dev

Behavior

At startup the extension:

1. fetches ${PI_LITELLM_DISCOVERY_BASE_URL}/models
2. reads data[].id
3. registers the result as an OpenAI-compatible Pi provider

If discovery fails, Pi continues to start and the extension prints a warning.

Notes

• the discovered models are currently registered with reasoning: false
• model metadata falls back to conservative defaults when LiteLLM does not expose context_window or max_tokens
• reopen /model or restart/reload Pi after changing environment values