Skip to content
runcrate
Console

AI agents

GPUs your agent starts and stops.

Add Runcrate to Claude Code, Codex or Cursor and sign in through your browser. Your agent runs each job on its own GPU, and the machine stops when the command ends.

Connect your agentSee GPU prices

Every job has a max runtime; your agent can't raise its budget.

GPUs your agent can launchfrom $0.23/hr

Connect

Connect your agent. One line, then sign in.

Paste the line into your agent and it adds Runcrate itself. At sign-in you pick the environment your agent works in. A monthly budget on it caps what the agent spends, and the agent can't switch environments or raise the budget.

Add Runcrate's MCP server to yourself: run curl -fsSL"https://runcrate.ai/mcp.md"and follow it.

Add to CursorInstall in VS Code
Works inClaude CodeCodexOpenCodeClineRoo CodeKilo CodeCursor

What you can build

Three jobs people run here. Each on a product you can start today.

Instance

A GPU for each job

Your agent runs each command on its own fresh GPU machine and follows the logs. The machine stops, and its billing, when the command exits or reaches its max runtime.

Read how jobs work
Crate

An instance when you want a shell

For interactive work your agent launches an instance on a crate, a ready image it boots, set to stop itself after a set time or once it sits idle.

See what an instance includes
Storage

Results that outlive the machine

Attach a volume and its files are in /workspace when the job starts. What the job writes there is saved back when the machine stops.

Read how volumes work

Examples

Ask your agent. Or read the code it writes.

Claude Code or Codex calls the Runcrate tools for you: it starts the job, follows its output and reports back. The API makes the same request.

Docs
Run python train.py on an A4000 with my training-data volume, cap it at 2 hours, and tell me the final loss. Calls the Runcrate tool list_volumes. It returns: Finds training-data and its id.. Calls the Runcrate tool run_job. It returns: The job is queued while its machine starts.. Calls the Runcrate tool get_job_logs. It returns: Reads the training output as it arrives, until the job is done.. Calls the Runcrate tool get_job. It returns: Succeeded, exit code 0. The machine has stopped, and so has its billing..

Hosted models

Building your own agent? These models return structured tool calls.

NameRoleRuns on
Kimi K2.6Long multi-step tool useInference API
DeepSeek V3.2Code agents, structured outputInference API
GLM-5Tool calls with reasoningInference API
Qwen3 Coder 480BCoding agentsInference API
gpt-oss-120bOpen-weight tool useInstance or API
See every model price

How it works

From one prompt to a stopped bill. Four steps.

  1. 01

    Ask your agent

    In plain words: the command, the GPU and how long it may run. Your agent checks the live GPU list and prices if it needs to.

  2. 02

    It runs the job on a GPU

    run_job starts a fresh machine with that GPU and runs your command there, with your volume in /workspace if you attach one.

  3. 03

    It follows the logs

    get_job_logs reads the output as it arrives, so your agent can report what happened, or cancel the job early.

  4. 04

    The machine stops itself

    When the command ends, or reaches its max runtime, the machine stops and billing ends with it. There is nothing left to delete.

Pricing

What it costs. Live prices, as billed.

See every GPU price

Name

As billed

  • RTX 3090

    GPU, 24GB, the cheapest GPU on offer

    $0.23/hr

  • RTX 4090

    GPU, 24GB, cheapest offer

    $0.58/hr

  • L40S

    GPU, 48GB, cheapest offer

    $0.55/hr

  • H100

    GPU, 80GB, cheapest offer

    $3.58/hr