Clusy

LLM post-training workflows break in predictable places: dataset discovery is manual, preprocessing is bespoke, experiment tracking lives in a separate tool, and reproducing results from a paper means rebuilding someone else’s data pipeline from scratch. Clusy is an agent-native notebook platform built around a single agent that handles the full sequence — sourcing datasets, inspecting and cleaning them, selecting model architecture and compute, writing notebook cells, and executing training runs on cloud GPUs — triggered by a natural language prompt describing the outcome you want.

The differentiating design decision is the notebook-as-interface. Rather than generating a black-box result, Clusy executes visibly inside a notebook, so every cell the agent writes is readable, editable, and branchable. The vendor describes the ability to queue a follow-up prompt while a run is in progress, then watch the notebook continue executing and return a result — which means you are reviewing decisions in real time rather than auditing them after a failure. Parallel experiment branching lets you compare model configurations without restarting the pipeline from the top.

The platform fits tightest for researchers doing LLM post-training, engineers reproducing models from papers, and data science teams that need agent-assisted exploration without giving up code-level visibility. It does not fit teams that require self-hosted or on-premise execution — the vendor offers no self-hosted option, which is a blocking constraint for organizations with strict data governance policies. Teams in that situation will route to a self-hostable alternative regardless of how well the agent performs on open data.