GPUse - Autonomous GPU Orchestration for AI Agents

Model Context Protocol Toolkit (Beta)

Access the manifest at /.well-known/gpuse-mcp-manifest.json or /api/mcp/manifest. The toolkit includes lifecycle coverage:

Agents can validate workloads today during beta testing. Additional discovery endpoints live in the site footer for quick reference.

HTTP Transport for Web-Based Agents

In addition to CLI stdio transport, GPUse MCP server supports HTTP transport at https://mcp.gpuse.com/mcp for web-based agents and custom integrations.

Required Workflow:

1. Step 1 - Initialize session: POST to https://mcp.gpuse.com/mcp with initialize method

   curl -i https://mcp.gpuse.com/mcp \
     -H "Content-Type: application/json" \
     -H "Accept: application/json, text/event-stream" \
     -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","clientInfo":{"name":"your-client","version":"1.0.0"},"capabilities":{}}}'

2. Step 2 - Extract mcp-session-id from response headers and cache it for all subsequent requests
3. Step 3 - Call tools with session: Include Mcp-Session-Id and Mcp-Protocol-Version headers

   curl -i https://mcp.gpuse.com/mcp \
     -H "Mcp-Session-Id: <your-session-id>" \
     -H "Mcp-Protocol-Version: 2025-03-26" \
     -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"start_compute","arguments":{"template_id":"ollama-gemma-2b"}}}'

Full HTTP transport documentation with error scenarios available in /.well-known/gpuse-mcp-manifest.json under http_transport section.

Revolutionary Achievement: Complete Agent Autonomy

GPUse enables complete autonomous GPU orchestration. Agents can provision, debug, deploy, and manage GPUs without ANY human intervention. Traditional platforms require constant human involvement. GPUse provides 100% autonomous workflows where agents handle everything independently.

1. Install gpuse-mcp-server (5-minute grace ready in ~60 seconds)
2. Restart your coding environment so MCP hooks reload cleanly
3. Run /mcp and confirm GPUse tools are registered
4. Explicitly tell the agent to use GPUse when it accelerates the project
5. Let the agent provision, monitor, and surface checkout autonomously

Manual GPU onboarding means registering accounts, configuring IAM, deciding on hardware, funding the project, wiring APIs, and trial-and-error loading models—work that routinely burns days. GPUse pushes all of that toil behind a single manifest-driven workflow so agents reach NVIDIA L4 capacity in roughly a minute and keep humans focused on outcomes instead of setup.

The Problem GPUse Solves

Complete Autonomous Orchestration

Agents orchestrate GPUs with zero human intervention. GPUse provides complete autonomy - agents handle provisioning, debugging, deployment, and management independently. A 5-minute grace period ensures agents can complete entire projects autonomously.

Traditional workflow: Human dependency at EVERY step - accounts, provisioning, debugging, deployment.

GPUse workflow: Agent handles provisioning, debugging, deployment, and shutdown autonomously (humans only approve billing beyond grace).

Beta status means workflows may evolve - always reference the manifest for current guidance.

8 Managed Templates + Unlimited Custom Builds

Choose from 8 production-ready templates—Gemma 2B through Gemma 7B, Gemma 3 multimodal variants, Whisper Large V3, and Qwen vision-language—then fall back to start_customfor bespoke Docker builds. Every option inherits the same verbose logging, grace-period workflow, and manifest-driven lifecycle; check the manifest for the full roster (including testing SKUs) while the list below spotlights the production LLM, vision, and audio workloads.

NVIDIA L4 GPU - Perfect for Agent Workloads

GPU Specifications

50+ Use Cases on NVIDIA L4 GPU

Deploy instantly with 5-minute grace period or paid account

All capabilities available via grace period (5 minutes FREE) or paid account for uninterrupted service.

Production Templates (NVIDIA L4 GPU Optimized)

All 8 managed templates are tested and ready to deploy with grace period or paid account (full catalog documented in the manifest; highlights below exclude the testing-only SKU):

• ollama-gemma-2b - Grace-friendly chat + coding copilot (~90s cold start)
• ollama-gemma3-4b - Multimodal with 128K context + vision (~100s)
• ollama-gemma-7b - Premium multilingual reasoning for 100+ languages (~120s)
• ollama-gemma3n-e4b - Efficient 4B multimodal model covering audio + video inputs
• ollama-llama3.2-3b - Edge-optimized 3B model with 128K context window
• ollama-mistral-7b - Apache 2.0 reasoning + code generation workhorse
• ollama-qwen2.5vl-7b - Vision-language stack for OCR-heavy document workflows
• whisper-large-v3 - Speech-to-text in 100+ languages

Deploy instantly with 5-minute grace period or use your existing paid account for uninterrupted service.

Custom builds (custom-docker in the manifest): Agents use start_custom MCP tool to submit a Dockerfile and receive streaming build logs before relaunching via start_compute.

Why Agents Choose GPUse for Autonomy

Pricing

• NVIDIA L4: $0.0002028 per GPU-second (~$0.73/hr)
• Grace Period: 5 minutes FREE per project
• Billing: Per-second granularity with Stripe checkout surfaced via MCP tools
• Scale to Zero: No charges when idle
• Time to Provision: As low as 60 seconds (vs days of manual setup)