# Krauncher

> Krauncher is a task-level GPU compute service, not a GPU provider. For each
> Python task it predicts the compute the task needs, estimates the real cost of
> running it at current market prices across multiple cloud GPU providers, and
> runs it on the cheapest suitable GPU. You don't pick a provider or rent a
> machine by the hour — you get an upfront estimate, and the bill is the GPU
> time the task actually used plus a flat dispatch fee, not a fixed price per
> finished task. Developer experience: add a decorator, await your function,
> get the result.

## Try it without an account

Krauncher prices a task before it runs, and that estimate is open — no key, no
signup. Paste code into the estimator on https://krauncher.com, or call it:

```
POST https://krauncher.com/api/estimate
  X-API-Key: cas_...        # optional — anonymous callers get an IP quota
  {"code": "<source containing a @client.task function>"}
```

It returns every eligible GPU with predicted time and cost, cheapest first,
priced against the live cross-provider market. Nothing executes and no GPU
time is spent, so it is safe to call in an optimization loop. Full response
shape in AGENTS.md.

For agents, the same pre-run estimate is also an MCP tool: `pip install
krauncher-mcp` — one `estimate` tool, no key (IP-quota'd). See the client list
below.

## Python client library

- [krauncher on GitHub](https://github.com/Ilya-a-sergeyev-ger/krauncher): the official Python client library and source of truth for the API.
- [README](https://github.com/Ilya-a-sergeyev-ger/krauncher/blob/main/README.md): install and quickstart.
- [AGENTS.md](https://krauncher.com/AGENTS.md): dense, accurate API reference for LLMs and coding agents — canonical pattern, `@client.task` parameters, TaskHandle/TaskResult fields, the values API (`run_code`), analysis as a phase separate from execution (`estimate_code` → reuse the classification), the `%%krauncher` cell magic and its flags, task groups, files in / artifacts out, storage credentials sealed to the worker, volumes and registered data sources, the pre-run `/estimate` endpoint, exceptions, constraints.
- [Examples](https://github.com/Ilya-a-sergeyev-ger/krauncher/tree/main/tutorial): numbered, runnable tutorials.
- [krauncher-jupyter on GitHub](https://github.com/Ilya-a-sergeyev-ger/cas-jupyter): the `%%krauncher` Jupyter cell magic — run a marked notebook cell on a remote GPU.
- [krauncher-mcp on PyPI](https://pypi.org/project/krauncher-mcp/): an MCP server exposing the pre-run estimate as one agent tool, `estimate` — `pip install krauncher-mcp`, works keyless (no account, IP-quota'd), for MCP clients like Claude and Cursor. Returns the code's cost profile on the reference card; see AGENTS.md.

## Site

Open to anyone, no account:

- [Landing & estimator](https://krauncher.com): what the service does, plus the live estimator described above.
- [GPU catalogue & live prices](https://krauncher.com/gpus): every GPU Krauncher can rent — VRAM, architecture, FP16/FP32 TFLOPS, memory bandwidth — with the current hourly price from each provider side by side, filterable by data region.
- [Pricing model](https://krauncher.com/pricing): how a task is priced — predicted compute against current multi-provider GPU market prices, and what the flat dispatch fee is.
- [SLA](https://krauncher.com/sla): draft failure & compensation policy — who pays when a run produces no usable result.
- [Privacy policy](https://krauncher.com/datenschutz): data processed, purposes, processors, storage periods (DE and EN).

Behind a login — an anonymous fetch of these returns the sign-in page:

- [Web app](https://krauncher.com/dashboard): dashboard, playground, tasks, storage and data sources, billing, API keys.
- [Python Library page](https://krauncher.com/library): install and quickstart on the site.