Skip to content
runcrate
Console
Blog

What we shipped since October 19

Clusters, jobs, model observability, secrets, a new playground and more. Everything you can do on Runcrate today that you could not do a week ago.

Jish

We've had a blog on the to-do list since roughly forever. It kept losing to shipping things, which is a nice problem, but it meant nobody knew what we shipped. So, sorry, and hello.

On October 19 we set ourselves one goal: make Runcrate the place your team and your agents go for GPUs and models, and make every part of it something you can use without asking us. This is what changed since then, grouped by what you can now do.

Clusters

  • Self-serve GPU clusters. Start 2 to 64 GPU machines on one private network, with Slurm, Kubernetes, Ray or nothing installed, from the dashboard, your agent, the CLI or the API. A cluster starts whole or not at all, and billing starts only when every machine has joined. Clusters docs
  • Priced by the machine. Every offer shows the whole cluster's price per hour and each machine's under it, and only sizes the location can deliver are offered. Open Clusters

Machines and launch

  • A failed launch is retried for you. When a machine fails before it boots, we start another on the same GPU, in the same place, at the same price or lower. When there is none, we ask you to pick, and you get one email either way.
  • Auto stop. Give any instance a maximum runtime or an idle stop, at launch or while it runs, so a forgotten machine stops on its own. Guardrails
  • Start later. Schedules start an instance or a job at a time you pick, once or on a cron. Schedules
  • HTTPS links and a firewall. Open a port on an instance as an HTTPS link for your workspace, anyone with the link, or a list of IPs, and set inbound rules per environment. HTTPS links
  • A real terminal in the browser, with an SSH fallback while the web terminal starts.
  • Prepaid terms. Prepay 1, 3, 6 or 12 months when you launch an instance on a machine that offers it, and renew from 7 days before the term ends. Without a renewal, the instance is terminated when the term ends.
  • A GPU picker with search, and machines with 1, 2, 4 or 8 GPUs. See GPUs

Jobs

  • Run one command on its own GPU machine. Runcrate starts the machine, streams the logs, keeps your outputs for 30 days and stops the machine when the command exits. Jobs wait for capacity across a list of GPUs, retry on a failed machine, run as arrays and sweeps, and take your code as an upload or from git. Jobs docs
  • A Jobs page that follows your first job step by step, then lists every run with its logs and outputs. Open Jobs

Models and inference

  • One playground for every model: chat, image, video and speech, with what each run cost and took, saved runs, and the code for the call. Open the playground
  • Claude Code and Codex on Runcrate models. The API now speaks Anthropic Messages and OpenAI Responses as well as Chat Completions. Anthropic Messages
  • A free tier for coding agents. Sign in with GitHub and open models are free within an allowance that resets on its own. How it works
  • Spend limits per API key, and no request-rate cap on paid models. Give a key a monthly spend limit and, if you want one, a request rate, checked before a request runs, so a call past a limit is refused and costs nothing. API keys
  • Every model with its maker's mark and verified capabilities, so you can see which ones call tools, read images or reason before you pick one. Embedding models are back too. Model catalog

Observability

  • Every model call, kept 30 days. How long it took, time to first token, speed, tokens, why it stopped and what it cost, filtered by model, key, end user, session or tag, and downloadable as CSV. Never the prompt or the answer unless you turn that on. Observability docs
  • Alerts when calls fail, by email and webhook, at a share you pick.
  • Send your calls to Langfuse, Datadog, OpenTelemetry, PostHog or a webhook, and score answers with feedback from code, an agent or the dashboard. Integrations

Storage and secrets

  • Secrets per environment, encrypted, write-only and pulled by the machine at boot, so a value never sits in a launch script. Secrets
  • Doppler, Infisical and 1Password, linked to an environment and read fresh at every launch. Open Integrations
  • Your files are copied to the volume one last time before an instance is deleted. Storage

Billing

  • A monthly budget per environment. You get an email and a webhook at 80%. At 100% new launches are refused, the environment's running instances are terminated and its API keys stop working until the next month or a higher budget. Budgets
  • Pay by bank transfer, with an invoice, when you add more credit than a card allows. Open Billing
  • A declined top-up says why, and an auto-recharge your bank asks you to confirm waits for you instead of failing.

API, CLI and MCP

  • Connect your agent with one line or one click in every major MCP client, signed in through the browser with a token you can revoke. MCP setup
  • The CLI runs jobs and clusters: runcrate run, runcrate jobs and runcrate clusters. CLI
  • Webhooks rebuilt, with Slack and Discord, signed deliveries, a test send, retries and a delivery log. Webhooks
  • Instances report ssh_host and ssh_port in the API, so a script can connect without parsing anything.

That is the week since October 19. We will write here whenever something you can use ships.

Subscribe

Get the next post by email.

One email per post. Unsubscribe in one click.

OlderJobs: one command on its own GPUOctober 26, 2026