Writing
Blog
How we build Clusy — releases, engineering write-ups, and the things that broke on the way. We post when there is something you can do that you could not do the day before.
The Benchmark Gap: Coding Agents vs Data Science
Coding agents got very good partly because we built a strong execution and verification loop around the model. For data science, we are still figuring out what that loop should be.
Math has Lean. Computational research needs a trace.
OpenAI just released 722 math manuscripts written by a model, and backed them with Lean proofs. Computational science has no proof checker. The closest thing it can have is a record of what actually ran, complete enough that anyone can run it again.
Data Science as a Search Problem
The next experiment depends on the last result. Branchable notebook state gives data science agents more than one path to explore.
Data Science Needs More Than a Coding Agent
The deeper problem is that data science isn't primarily a code generation problem.
Towards Fluid Compute - Making Python Sessions Portable
The machine is not the computation. It is just where the computation happens to be running right now.
Clusy has a terminal now
Install a package, check a dataset, follow a training log, or see the agent's shell commands without leaving your notebook.
Clusy now runs R natively
Pick R when you start a notebook and the agent writes the R, runs it, checks the plots, and gives you the finished notebook. A real R session, in the same workspace as your Python one.
Clusy just got better
A live-kernel agent, parallel experiments, and workspace snapshots that survive a bigger machine. Every plan also has higher limits, and everyone's credits are fresh.
Kimi K2.7 and K3, GLM 5.2 and Qwen3.8 Max, free for everyone for three days
Four open-weight flagships in the model menu — and through August 8 they cost nothing, on every account. Clusy ❤️ open source.
Qwen3.8 Max, free for everyone this week
Alibaba's 1M-token-context flagship joins the model menu — and now through July 27 it costs nothing, on every account.
Kimi K3, free for everyone this week
Moonshot's 1M-token-context flagship is in the model menu — and until July 23 it costs nothing, on every account.
Share a notebook someone can actually run
A Clusy link opens the whole notebook — cells, outputs, branches — for anyone, with no account, and they can fork it into their own workspace in one click.
The agent retracted its own result
We handed Clusy a client's raw email and a ZIP of camera frames and walked away. It came back with 293 auto-labelled frames, three problems in the data we hadn't noticed — and, after we pushed back, a withdrawal of its own headline finding.
The agent was labeling blind
Our labeling agent kept shipping confident, wrong boxes. The model was fine. Jupyter was silently deleting the images it used to check its own work.
Bring your own key, or your own machine
Connect an Anthropic or OpenAI API key and run Sonnet 5, Opus 4.8 or GPT-5.6 Sol for zero Clusy credits — or pair the Claude Code or Codex CLI you already have and let Clusy drive it on your own subscription.
We open-sourced our status page
A self-hosted status page for Next.js and Vercel — live health checks, 90-day uptime, incidents that open and close themselves. No database. MIT.
Introducing branched experiments
Describe a comparison and Clusy forks one branch of your notebook per variant, runs them side by side on their own kernels, and brings the results back as a single comparison.
Changelog
Every change we ship, in the order we ship it.