The open-source AI research agent,built on Claude Code & Codex.
THE RUNCELL TEAMCREATORS OF PYGWALKER
Abstract
We don't rebuild the agent harness. Runcell Science runs on Claude Code and Codex — the strongest harnesses that exist — and puts everything into the research workflow layer: sessions, artifacts, and skills. It is local-first and Apache-2.0 licensed, and your existing subscription already covers it.
Keywords: open-source AI research agent · Claude Code · Codex · research sessions · self-hosted
$ git clone https://github.com/runcell-ai/runcell-science.git▸ session started · harness: claude code · workspace: ~/lab/assay
youCompare the three candidate models on the new assay data and produce a summary figure.
agentReading data/assay_results.csv — 3 models × 240 runs. Writing analysis.py…
$ python analysis.py✓ exit 0 · 3.1s
agentModel B leads on both metrics; full table below the figure.
⬒ artifact · fig_model_comparison.png✓ checkpoint saved · turn 4
Why we refuse to rebuild the agent harness
Every agent product sits on a harness — the machinery that manages context, plans, calls tools, executes code, and recovers when things break. Our bet is that this layer is already solved, and not by us.
Harnesses are the hard part
Claude Code and Codex embody enormous engineering investment: context management, tool orchestration, sandboxed execution, failure recovery. A harness rebuilt from scratch starts years behind — and every hour spent rebuilding it is an hour not spent on research workflows.
Your subscription already covers it
Runcell Science drives the claude and codex CLIs you are already signed in to. No separate API keys, no second bill, no hosted proxy sitting between you and the model. If your plan runs Claude Code today, it runs Runcell Science today.
Effort goes where it differentiates
Our engineering concentrates on the research workflow layer: session management, artifacts, and skills for scientific work. When the underlying harnesses improve — and they improve monthly — Runcell Science gets better with zero migration.
“This is an opinionated position — and a verifiable one. The code is open; read it.”
A workspace for agent-assisted research
Research moves across code, data, papers, notebooks, and experiments. Runcell Science keeps that loop — and everything your agent produces along the way — in one local, inspectable place.
- 2.1
Sessions that hold the whole thread
Agent replies, tool activity, and generated work stay attached to one research session, with checkpoints recorded turn by turn. Long-running work stays traceable instead of scattering across terminals and chat tabs.
- 2.2
Claude Code & Codex, first-class
Both harnesses ship as native providers. Runcell Science talks to the local claude and codex CLIs you already have installed and signed in — the strongest agent runtimes available, not a reimplementation of them.
- 2.3
Artifacts stay with the work
Figures, notebooks, and files the agent produces are captured as artifacts on the session and rendered in place — not lost somewhere in a scroll of chat history.
- 2.4
Skills for research workflows
A skills layer packages repeatable research procedures for the agent. The library is early and growing — and growing it in the open is precisely the point of an open-source project.
- 2.5
Local-first, by design
Runs on your machine, in your browser, against your project files. Nothing routes through a hosted backend. Apache-2.0 licensed, self-hostable, auditable — read every line before you trust it.
- 2.6roadmap
Desktop builds, including Windows
Packaged desktop apps are planned, with Windows treated as a first-class target rather than an afterthought — an underserved gap across this entire category.
Three ways to build a research agent
Several open-source Claude Science alternatives launched within days of each other. They differ on one structural decision — the harness — and we think that decision determines the rest.
Table 1. Architectural comparison across the three approaches.
| Dimension | Runcell Science | Self-built harness agents | Claude Science (official) |
|---|---|---|---|
| Agent harness | Claude Code & Codex | Custom-built | Anthropic internal |
| Reuses your existing Claude Code / Codex subscription | Yes | No — separate API keys | No |
| Open source / self-hostable | Yes — Apache-2.0 | Yes | No |
| Cross-vendor (Claude + Codex) | Yes | Varies | No |
| Skills library size | Growing | Often larger | Large |
Note. Yes — others currently ship more skills. Skills are the easiest layer to grow in the open; the harness is the hardest to rebuild. That trade is the point.
First session in five minutes
$ git clone https://github.com/runcell-ai/runcell-science.git
$ cd runcell-science
$ ./scripts/dev.shhttp://127.0.0.1:27183$ yarn install
$ yarn devFrequently asked questions
Is there an open-source alternative to Claude Science?
Yes. Runcell Science is an open-source alternative to Claude Science, released under the Apache-2.0 license. It runs locally on your machine, keeps your research files under your control, and uses Claude Code or Codex as its agent harness — so it works with the subscription you already have.
Does Runcell Science work with my existing Claude Code subscription?
Yes. Runcell Science drives the Claude Code CLI you already have installed and signed in. If your Claude plan covers Claude Code, it covers Runcell Science sessions too. The same applies to Codex and your existing OpenAI plan.
Do I need my own API keys?
No. Runcell Science does not proxy your requests through a hosted service or bill you per token. It runs on top of the local claude or codex CLI, authenticated with the account you are already signed in to, so there is no separate API-key billing.
How is this different from building an agent with its own harness?
The harness — the layer that manages context, orchestrates tools, executes code safely, and recovers from errors — is the hardest part of an agent to get right. Projects that rebuild it from scratch start from behind and usually require separate API-key billing. Runcell Science builds on Claude Code and Codex instead, and spends its effort on the research layer: sessions, artifacts, and skills.
Is Windows supported?
A native Windows desktop build is on the roadmap — it is a gap in this category we intend to fill. Today Runcell Science runs as a local web app; on Windows, running it under WSL is the practical path. Track progress in the GitHub repository.
How does Runcell Science relate to Runcell?
Runcell Science is an open-source project by the Runcell team, the creators of PygWalker. Runcell itself is a separate commercial product — an AI agent for Jupyter notebooks. Runcell Science is independent and Apache-2.0 licensed; you do not need a Runcell account to use it.