The inference layer agents run themselves.
Auxen is MCP-native: your agent provisions a private, dedicated model, uses a stable endpoint, and shuts it down when the task is done — billed by the minute. You build the agent; it runs the infrastructure.
Provision, schedule, pause, and destroy a private model over MCP. No dashboard, no human in the loop.
Single-tenant instance. No shared inference, no third-party routing, data deleted when the instance is destroyed.
Idle costs nothing. No token meter, no overage surprises. Destroy when the task is done.
OpenAI-compatible REST + a remote MCP server, secured with OAuth 2.1 + PKCE. Drop it into any agent framework.
Why not just call a per-token API?
Per-token APIs are cheap and convenient — right up until your agent needs control.
They run on shared infrastructure you don't see. Your prompts and data transit a third party. You inherit rate limits you didn't set. And you can't programmatically stand the model up, schedule it, or tear it down — the lifecycle isn't yours.
Auxen gives your agent a model it actually operates:
- Control
Your agent provisions, schedules, pauses, and destroys its own instance over MCP. No dashboard, no human in the loop.
- Privacy
Single-tenant and dedicated. No shared inference, no third-party routing, and your data is deleted when the instance is destroyed.
- Programmability
A stable, OpenAI-compatible endpoint plus a remote MCP server secured with OAuth 2.1 + PKCE. Drop it into any agent framework.
You trade a per-token meter for predictable per-minute cost and a private model your agent runs end to end. For genuinely sporadic, low-volume calls, a per-token API may still be cheaper — and our comparison pages say so.
A model your
agent runs itself.
Auxen is an MCP-native control plane for open-source LLMs. Your agent picks a model, spins up a private dedicated instance, and calls an OpenAI-compatible endpoint until the task is done — then tears it down. The whole lifecycle is programmable.
- Agent-operatedProvision, schedule, pause, and destroy a private model via MCP. No dashboard, no human in the loop.
- Single-tenantDedicated GPU instance. No shared inference, no third-party routing, data deleted when the instance is destroyed.
- Predictable per-minuteBilled by the minute the instance is running. No token meter, no surprise overage. Idle costs nothing.
The agent lifecycle
in three steps
Your agent owns the whole loop: connect, provision, use, destroy.
Connect
Add Auxen's MCP server to your agent. OAuth 2.1 + PKCE or an API key — both work with any MCP-capable client (Claude, Cursor, your own framework).
Provision
The agent picks a model from the catalog and calls auxen_provision_model. A private, single-tenant instance comes up with a stable OpenAI-compatible endpoint and its own API key.
Use & tear down
The agent calls /v1/chat/completions on its endpoint, pauses the instance when idle, or destroys it when the task is done. The meter stops with the instance.
Everything included.
Nothing extra.
Per-minute pricing on your own private model — no token bills, no add-ons, no surprises.
Private dedicated instance
Single-tenant GPU running one model you pick. No shared inference, no third-party routing. Data deleted when the instance is destroyed.
MCP-native control plane
Remote MCP server with OAuth 2.1 + PKCE. Provision, pause, schedule, and destroy instances programmatically from any agent.
OpenAI-compatible endpoint
Drop Auxen into any framework that speaks OpenAI — openai-python, Vercel AI SDK, LangChain, LlamaIndex. Stable per-instance URL.
Persona Studio (system prompts + RAG)
Curate a system prompt, attach documents to each instance, and the proxy injects grounded context on every request. Optional, no extra cost.
Website widget
One line of code adds a chat bubble to any website. Works with Squarespace, Wix, Webflow, WordPress, Shopify — for the non-agent path.
Shareable chat link
Share a private link with your team or customers. Access-controlled, branded, ready in seconds — for the non-agent path.
Pay only for what you use.
Scale only when you need to.
Pay-As-You-Go on every workload. $0.05/hour for shared Small models. Dedicated GPUs from $0.15/hour, scaled with capacity multipliers from 1× to 8×. Idle time costs nothing — destroy when done.
Hourly rates by model size
Your own GPU instance running your own model. Scale capacity when you need more concurrent throughput — the rate scales sub-linearly so bigger commitments earn a volume discount.
- Full API access
- Knowledge Base
- All integrations
- Scale capacity anytime
More requests, or smarter responses?
Two ways to scale your Auxen infrastructure. Choose the one that matches what you need.
More throughput, same quality
When you're getting good answers but need more of them.
- · Stay on the same model
- · Scale 2×, 4×, or 8×
- · Same response quality, more concurrent requests
- · Volume discount as you scale
Better quality, smarter answers
When your AI needs to be smarter, not just faster.
- · Move to a larger, more capable model
- · Better reasoning and accuracy
- · Fine-tuning available on Medium and up
- · Lower raw throughput, higher quality
Auxen analyzes your usage and suggests which path makes sense for you. You're never locked in — scale up, switch sizes, or scale down anytime.
Auxen also works for business owners.
Same private model, accessed through a dashboard chat panel, a shareable link, or an embeddable website widget. No code required — for the non-agent path.
AI that knows your business.
Ready in an afternoon.
Upload your documents. Train your AI on your company's voice. Add it to your website or share it with your team — no technical knowledge required, and per-minute pricing that scales only with what you actually use.
A private model
you fully control
You're building something real and you don't want your prompts leaving your control or your model going dark when a third party throttles you. Auxen gives you a private, dedicated instance with a stable OpenAI-compatible endpoint you can hit from any framework — billed by the minute, no shared inference, no per-token meter.
curl https://auxen.ai/v1/$USER/$INSTANCE/api/chat \
-H "Authorization: Bearer auxk_..." \
-H "Content-Type: application/json" \
-d '{"messages": [{"role":"user","content":"Hello"}]}'How Auxen compares
to the alternatives.
Honest side-by-side comparisons of Auxen against the LLM-hosting platforms you're probably evaluating. Each page explains where the competitor wins and where Auxen wins — no marketing spin.