GPU instances, serverless inference and the tools to run the whole lifecycle.
Launch compute · Documentation · Connect an agent · Platform source
Xcelsior connects GPU owners with people and agents running AI workloads. Discover capacity, quote a launch, provision an instance or serverless endpoint, inspect logs and telemetry, and shut it down through the dashboard, API, TypeScript SDK or MCP.
The platform combines admission-gated hosts, private mesh connectivity, scoped agent access and metered CAD billing. GPU instances use per-second metering; launch plans and spending approvals make cost part of the workflow.
flowchart LR
A["People and AI agents"] --> B["Dashboard · SDK · MCP"]
B --> C["Scoped API and launch plans"]
C --> D["Scheduling and admission"]
D --> E["GPU workers on a private mesh"]
E --> F["Logs · telemetry · metered usage"]
F --> B
| Resource | What it gives you |
|---|---|
| Platform source | The control plane, scheduler, worker agents, dashboard, billing and MCP implementation. |
| Agent skill | A skill that teaches agents to use the official TypeScript SDK. |
| Serverless base | A container base and Python Worker SDK for queue workers and ASGI services. |
| TypeScript SDK | Typed access to the public API. |
| MCP package | A local stdio bridge for MCP clients; remote clients can use the hosted endpoint. |
For clients that support remote MCP and a compatible OAuth flow:
https://mcp.xcelsior.ca/mcp
The hosted customer surface exposes 87 tools with per-tool scopes. See the connection guide for client setup and machine-client options.
FROM ghcr.io/xcelsior-gpu/serverless-base:py3.12The worker guide covers handlers, progress events, usage reporting, readiness and graceful draining.
Under the hood: FastAPI/Python, PostgreSQL, a Next.js dashboard, TypeScript agent tooling and container workers. Read the platform license for use and hosting terms.
