Verse

Verse is the lab’s OpenAI-compatible inference gateway. Apps and agents call one endpoint and Verse routes to local GPU models on aistack or to hosted fallbacks, so clients never hold the upstream provider keys.

Used by Jello (Pi harness), mem0 and DeerFlow.

User Guide

Models

ModelCapabilityBackend
Qwen/Qwen3.8-27B-FP8chatlocal, aistack
Qwen/Qwen-Image-2.1-Turboimageslocal, aistack
muse-spark-1.3-contributorchatOpenCode Go
gemini-3.5-flash-litechatGemini
gemini-embedding-001embeddingsGemini

GET /v1/models lists what’s live.

How to Connect

Needs the lab network (Coder or Wireguard) and the GGL CA, see Trusting Vault root CA.

Routes: /v1/models, /v1/chat/completions, /v1/responses, /v1/embeddings, /v1/images/generations, /v1/images/edits.

curl https://verse.k8s.internal.galaxygridlabs.com/v1/chat/completions \
  -H "Authorization: Bearer $VERSE_API_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"model":"Qwen/Qwen3.8-27B-FP8","messages":[{"role":"user","content":"hi"}]}'

How to Get an API Token

Login is Vault OIDC, restricted to labusers@hul.to and labadmins@hul.to. Tokens are minted over OAuth from the verse service dir:

curl -fsS https://vault.galaxygridlabs.com/v1/pki2/ca/pem -o /tmp/ggl-ca.pem
SSL_CERT_FILE=/tmp/ggl-ca.pem python3 scripts/demo-oauth.py \
  --url https://verse.k8s.internal.galaxygridlabs.com \
  --scenario token --token-name <name> --token-file /tmp/verse-token

Open the printed URL in a browser that can reach the loopback callback (:8765).

Note 📝 - The token (vrs_...) is shown once. Max 20 active tokens per user, each expiring after 90 days. Manage them at /v1/api-tokens (GET list, DELETE /{id} revoke). API tokens can call inference but can’t manage tokens.

MCP / OAuth clients

Verse is also an OAuth authorization server (dynamic client registration, Vault login), so OAuth-capable clients don’t need a token.

Admin Guide

Runs on smoll-harvester from src/fleet_smoll_harvester/verse with Postgres. Pulumi pushes the secrets and config in src/2_app/services/verse/verse.py.

Providers

Models and their upstream credentials are declared in the _config in verse.py. Upstream keys live in the Verse Secret and never reach clients. The local models are served by aistack vLLM, defined in src/2_app/services/aistack/aistack.py.

Static keys

VERSE_API_KEYS (comma separated, 32+ chars) in the Fleet Deployment gives service-to-service inference access, currently one for DeerFlow. They can’t manage tokens. Rotate by editing and restarting.

Limits

Request bodies 10 MiB, requests time out at 10m.

  • Jello
  • mem0
  • src/2_app/services/verse/API_TOKENS.md