Works with
vLLM
SGLang
NVIDIA NIM
Dell Enterprise Hub
The situation
Your gateway knows who is asking but sees your GPUs as just a URL. Your GPU stack knows the hardware but not the person. When a coding agent fails on a model you already run, nobody can say why.
Shared fairly, and green only when real work succeeds
First come, first served
Teams from your SSO groups get a share of each endpoint, and interactive requests go before background agents
Requests fail with errors
Background requests wait in a queue with a clear message
Endpoint keys handed to every team
The gateway is the only way in, the endpoint key never leaves it, and teams never share a prompt cache
One bill, and nobody knows who used what
Split by each team’s real token use, with idle cost shown separately
Green means the URL answers
A real agent turn with a tool call runs on each endpoint every few minutes, on the path your users take
Errors scattered across servers
One screen per endpoint, with recent failures, the engine’s error text, slow calls, and the conversation that failed
Found when users complain
Endpoint down, slow first token, long queue, full KV cache, or a team near its share, sent to email, Slack, or a webhook
Switch and hope
Past agent tasks replay on the candidate model first; the alias moves only after the replay passes, and rollback is one click
From an endpoint URL to agent-ready, in minutes
01
Connect
Paste an endpoint URL. xpander reads the models, engine version, context length, and serving settings, and checks again on a schedule.
02
Test
xpander replays real Claude Code, Codex, OpenCode, and xpander agent traffic on your endpoint, including tool calls, reasoning settings, and long contexts.
03
Fix
Every result is a pass, or a fail with the cause and the fix: a server setting to change, or a gateway fix xpander turns on.
04
Size
xpander measures time to first token, tokens per second, and how many people the endpoint can serve, then suggests limits per team.
Your endpoints next to every other model
Claude Code, Codex, and OpenCode on your own models
Your people use the coding tools they already know, on your GPUs, with their company login. Teams get your models with the same groups, limits, and keys as cloud models.
The gateway adapts each request to your engine: reasoning effort per model family, system messages where the model’s template accepts them, and context-full errors in each tool’s own format, so Claude Code compacts instead of failing.
Agent sessions on your models, in one call
Developers start an agent session on a self-hosted model with the same API shape as the OpenAI Agents API. They point the OpenAI SDK at xpander with an xpander key, and need no other client.
Pick the harness per session: Codex, Claude Code, OpenCode, or the xpander runtime. xpander runs it and keeps the session, with the same identity, limits, logs, and cost as chat, in the cloud or air-gapped.
AI for everyone, without your data leaving
Gateways, agents, the AI Workspace, and every log run in your cloud account or your data center, with your own models next to providers you approve, behind one gateway with the same access rules.
Fully air-gapped when required, with TLS to your endpoints on your private CA and only the gateway allowed to reach the GPUs.
Questions we hear from IT leaders
Any vLLM or SGLang endpoint, with NVIDIA NIM and Dell Enterprise Hub as named providers. xpander reads each engine’s own endpoints to find its models and settings.
