LiteLLM per-member keys

This zero-dependency Node.js example reconciles OpenWork Cloud per-member LLM credential bindings with LiteLLM virtual keys. It is an example provisioner, not a native LiteLLM integration in Den.

Configure

Set these environment variables:

OPENWORK_DEN_API_URL: Den API base URL
OPENWORK_DEN_TOKEN: organization owner or admin bearer token
OPENWORK_ORG_ID: organization ID sent as x-openwork-org-id
OPENWORK_LLM_PROVIDER_ID: the per-member LLM provider ID
LITELLM_BASE_URL: LiteLLM base URL, with or without /v1
LITELLM_MASTER_KEY: LiteLLM master key
LITELLM_MODELS: comma-separated model IDs assigned to each virtual key

Run reconciliation after granting provider access:

node provision.mjs reconcile

Before provisioning keys, the script reads the existing custom, per_member provider from Den and reconciles its configured model metadata from LiteLLM GET /model_group/info using master-key authentication. It preserves the provider config, model names and unknown model fields, and current member/team access while updating token limits. When LiteLLM supplies the corresponding facts, it also maps function calling, reasoning, vision, response schemas, and temperature support to Den's tool_call, reasoning, attachment, structured_output, and temperature model fields. It only PATCHes Den when those values changed.

Every model in LITELLM_MODELS must have an exact model_group match with finite, positive max_input_tokens and max_output_tokens. Reconciliation fails closed before creating member keys if LiteLLM omits a requested model or either limit. The example never guesses token limits or falls back to a generic value.

After metadata is synchronized, the script lists Den member credential states, mints a LiteLLM virtual key for each missing member, and writes the key to Den with LiteLLM's token_id as externalCredentialId. Summaries report the metadata action and safe model limits but never contain member keys.

Run the complete local world

From the repository root, launch a signed-in Desktop, an isolated Den organization, and a database-backed LiteLLM gateway with the provider and member keys already reconciled:

pnpm world up ./worlds/litellm-per-member.ts

The command prints the Den URLs, LiteLLM URL, synchronized model limits, Den provider record ID, and Desktop CDP URL. Keep it running while testing and press Ctrl-C to tear down the world. Docker, local MySQL, and local Redis are required. The gateway uses a deterministic local OpenAI-compatible witness and does not read or require OPENAI_API_KEY.

Offboard a member by organization membership ID:

node provision.mjs offboard member_...

Ordering rule: block the upstream LiteLLM key first, then mark the Den binding blocked. LiteLLM v1.97 accepts its generated token_id in POST /key/block, so offboarding does not need the plaintext member key.

See Per-member LLM credentials for the API contract. The executable proof is litellm-per-member-credentials.e2e.test.ts.