Documentation

Run obleth in front of your models, share them fairly, and see where every token goes. Written for v2.4.0.

Start here

5 minTry it on your laptopDocker Compose, a GPU-free test backend, and your first request.20 minDeploy on KubernetesThe Helm chart, your secrets, and pools sized from ready replicas.30 minRun models on SlurmConnect slurmrestd, pick a recipe, and launch a deployment.2 minPoint your clients at itOpenAI SDKs, Claude Code, and anything that speaks either.

Browse by task

ModelsRegister, import and keep them healthyRegister a model Pool sizes Health checks

Sharing capacityFair share, tenants, keys and budgetsTenants and groups API keys Quotas and rate limits

BoonsAbilities the gateway adds to a modelVision Tool loop Knowledge

Connect clientsEvery API the gateway speaksOpenAI-compatible API Claude Code and Anthropic SDKs Verdicts

DeploymentsModels on your own computeThe Deployments page Launch on Slurm Recipes

OperateRun it in productionProduction checklist Upgrades Troubleshooting

obleth-chat

A chat app for everyone who shares the gatewayPrivate conversations per person, document collections with citations, workspaces that build apps with a live preview, and your Slurm and Kubernetes clusters as tools.

New in v2.0

Full changelog

Fairshare across replicasPool limits are cluster-wide slots, so any replica can use all of them.Discovered pool sizesA model's pool follows its ready replicas instead of a typed number.Messages, Videos and VerdictsClaude Code against your own models, video jobs, and typed decisions.A redesigned dashboardEvery page leads with what needs you and saves with one bar.