Compute integrations
How to prepare your data and results for tools you already use, and what each guide does and doesn't connect for you.
Choose the integration level
Your agent keeps the model, goal, context, planning and final decision. Your application or a compatible supported client prepares the permitted input, while the Kit can help with the declared handoff where that workflow provides a Kit helper. Lithi accepts a supported workflow request or encrypted artifact reference after transfer permission, quote and approval, then returns the job or result identity for you to check.
Start with the data or result you need to move. These guides use three levels:
- File interchange: prepare a fixed export, then use the supported encrypted artifact handoff.
- Migration guide: map an existing provider request to a supported Compute workflow; its original envelope and endpoint do not carry over automatically.
- Actual adapter: use this label only when a guide names a supported native adapter. The guides below show customer-side preparation and local library use; they do not promise an automatic provider or framework adapter.
Translate the upstream task into a supported Compute operation before preparing its source. A dataset revision, JSONL file, table export or tracker record is local evidence for that request; it is not, by itself, a wire request or a reason to redirect an endpoint unchanged. The selected client or library carries the declared input and output contract, while access, approval and spend remain with the Compute account.
| System/tooling | Boundary level | Guide |
|---|---|---|
| Hugging Face datasets | File interchange | Hugging Face |
| Iceberg/Hudi | File interchange | Iceberg and Hudi |
| MLflow/W&B | File interchange for reviewed results | MLflow and W&B |
| On-premises sources | File interchange | On-premises |
| OpenAI Batch migration | Migration guide | OpenAI Batch |
| pandas/Polars/Arrow | File interchange | pandas, Polars and Arrow |
| PyTorch | File interchange for local loaders | PyTorch |
| Ray/Spark | File interchange for local frameworks | Ray and Spark |
None of these rows is an automatic provider or framework adapter. If you need an actual adapter, confirm that the selected workflow names one as supported before you prepare data.
A recognized format doesn't grant access to a source. A source connection doesn't grant permission to write to a destination.
For accounting systems, use Finance Sync and the Finance mappings reference.
For the supplied local modules and their exact helper boundaries, see Kit contents and the Kit helper reference.