Endpoints
Dedicated endpoints for your models.
Create your first endpoint.
Every endpoint has its own HTTPS URL, key and status.
Set up and run from your coding agent
Connect once, then let your agent quote, launch and use an endpoint. You approve the price before launch. Connecting costs nothing. Setup guide · REST API
See the prompt
Set up InferOptimal for me and continue through a working task.
1. Read https://www.inferoptimal.com/docs/agents.md (or inferoptimal://start when connected) and follow it.
2. Connect only if needed; the CLI is optional:
Claude Code: claude mcp add --transport http --scope user inferoptimal https://www.inferoptimal.com/mcp; then claude mcp login inferoptimal.
Codex CLI: codex mcp add inferoptimal --url https://www.inferoptimal.com/mcp; then codex mcp login inferoptimal.
Codex desktop: Settings → MCP servers → Add server inferoptimal (Streamable HTTP, same URL), then Authenticate.
Other MCP clients: {"mcpServers":{"inferoptimal":{"type":"http","url":"https://www.inferoptimal.com/mcp"}}}
I approve sign-in in my browser; never ask for a password or key in chat.
3. Use a suitable existing InferOptimal endpoint, or choose the offering for my task from query action list_models and quote it. Show the hourly price, runtime limit and the quote's usage plan before launching; use my existing approval for this quote, otherwise wait for my approval. Then configure it and run my task. Check query action list_dedicated_endpoints and query action list_deployments first to avoid duplicates. Reuse a suitable service and its saved key. If I name a model, use that model's offering. Do not substitute a different model.
4. Tell me the call you will make, run one real task, then continue at the quote's concurrency: the endpoint bills by the hour. Setup alone is not completion. Keep the endpoint running for my job. When finished, Stop the dedicated-key endpoint: it drains active requests, releases the hardware and ends billing; its URL and key stay the same. Start it again with the same key before the next job. Stop explicit deployments when finished.Set it up by hand
CLI alternative, if installed:
Connect a cloud
optional · InferOptimal cloud needs none
AWSnot connectedMore options
The stack creates one role. It can see which GPU instance types your account can launch, file quota requests with AWS, and launch, stop and terminate instances it tags velobyte=1; it cannot read your data or touch anything else. Delete the stack and every permission goes with it. Read the template.
Lambdano key
Modalno tokenModal connect is not available yet. InferOptimal has not turned on the service that deploys into your workspace, so no token is taken here. Until then you can make a deploy file and run it yourself.
Or make a deploy file and run it yourself
- In a terminal, run
pip install modal && modal setup. It opens modal.com, makes a token and keeps it on your computer. - Enter a model, pick the GPUs and make the file.
- Save the file and run the commands shown with it. The last one prints the endpoint's URL.
A file you deploy yourself runs on your computer's Modal token. InferOptimal cannot start, watch or stop it, and it does not appear under Endpoints. Modal bills your workspace, credits included.
Connect your project with short-lived credentials. No key file required.
Open Cloud Shell in your project and run this command. It creates a dedicated account with Compute Admin and Cloud Quotas Admin access. Then return here to verify.
