Runcell Science: An Open Source Alternative to Claude Science for Research Workflows

Name

Soren Atelier

Updated on 7/6/2026

If you are looking for an open source alternative to Claude Science, Runcell Science is worth watching. It is not just another chat interface for scientific questions. It is a local-first AI research workspace that brings agent sessions, local project files, open science connectors, tool activity, generated artifacts, code diffs, and follow-up work into one research environment.

The short version: Claude Science points toward a future where AI agents help with scientific work across papers, data, code, experiments, and research outputs. Runcell Science takes a more open, local-first route to that workflow. The workspace stays centered on your project, while the execution backend can use Codex, Claude Code, and future code agents.

That matters for developers, researchers, and research engineers because scientific work rarely fits inside a one-off prompt. You may need to scan literature, query a biological database, inspect a clinical trial, generate an analysis notebook, update a pipeline, review a diff, and turn the result into a reproducible artifact. Runcell Science is designed to keep those steps connected.

Project repo: https://github.com/runcell-ai/runcell-science (opens in a new tab)

Quick Take: Who Runcell Science Is For

In one sentence: Claude Science is best understood as a scientific AI product direction, while Runcell Science is an open source research workspace where Codex, Claude Code, connectors, local files, artifacts, and reviewable diffs can work together.

This also places Runcell Science in the broader open science tooling category. Open science is not only about open papers or open datasets. It also depends on workflows that make research more transparent, reproducible, inspectable, and easier to share. Runcell Science supports that direction through local project control, auditable agent sessions, explicit connectors, and diffs that can be reviewed before changes become part of the research project.

What Is Runcell Science?

Runcell Science is a local-first workspace for researchers, research engineers, students, and technical teams using AI coding agents in scientific work.

Instead of treating the agent as a separate chat window, Runcell Science treats the research project as the main object:

• agent conversations belong to concrete sessions;
• local files and project state become part of the working context;
• tool activity, generated outputs, artifacts, and diffs stay connected to the same workflow;
• follow-up tasks do not have to be recovered from a long chat history;
• connectors can be enabled or disabled for the current research session.

That structure is important because research work is iterative. A useful agent may need to move from literature review to notebook edits, from gene lookup to pipeline debugging, or from a prototype script to a report draft. Runcell Science is trying to make that process feel like one sustained workspace instead of a set of disconnected prompts.

Why It Is More Than a Scientific Chatbot

General AI chat is useful for quick explanations, but research development depends on continuity:

• Which repository, notebooks, and scripts are currently in scope?
• What did the agent change in the last pass?
• Which results came from PubMed, bioRxiv, Clinical Trials, ChEMBL, or another connector?
• Which artifact should be edited next?
• Which diff is ready for human review?
• Which connectors should this session expose to the agent?

Runcell Science is closer to an AI-assisted research development environment. You are not only asking for an answer. You are advancing a local research project with an agent that can read files, generate code, organize findings, produce artifacts, and leave reviewable changes behind.

Three Design Choices That Matter

1. Local-first project control

Runcell Science is positioned as a local-first workspace for using AI coding agents in research. Local-first does not mean the system never calls a model or external service. It means the project remains organized around your local workspace rather than being split into a hosted chat context.

That is a practical distinction for scientific and technical teams. Research projects often include:

• unpublished code and experimental ideas;
• local notebooks, scripts, configs, and intermediate files;
• long-running context that needs to survive beyond one chat;
• diffs that should be reviewed before they are accepted;
• reproducibility and audit requirements.

For research engineers, the benefit is simple: the agent can work on the real project instead of handing you snippets that still need to be copied back into the repo.

2. Replaceable agent backends: Codex and Claude Code

Runcell Science currently supports two agent providers in its runtime architecture: codex and claude. In practice, that means the workspace is not hardwired to one model provider or one agent experience.

Different research tasks can benefit from different agents:

This is one of the clearest differences between Runcell Science and a thin wrapper around a single agent. The workspace, connectors, sessions, artifacts, and diffs can remain stable while the execution backend changes according to cost, capability, privacy, speed, or team preference.

The backends are not interchangeable in every detail. Codex, Claude Code, and future agents may differ in tool protocols, permission models, context behavior, pricing, and execution style. The value of Runcell Science is that it gives research teams a place to manage those differences instead of forcing the entire workflow into one agent silo.

3. Connector-first research workflows

The strongest reason to care about Runcell Science is not that it adds a research-themed UI around a coding agent. The stronger idea is that scientific connectors are treated as first-class workflow infrastructure.

An AI research agent that can only chat and edit code is often missing the materials that scientific work depends on. Runcell Science includes a bundled science connector registry with connectors such as:

• BioMart
• PubMed
• bioRxiv / medRxiv
• Clinical Trials
• ChEMBL
• Genes & Ontologies
• Protein Annotation
• Structures & Interactions
• Variants
• Literature Graph
• Expression
• Omics Archives
• Regulation
• Drug Regulatory
• Research Resources
• Cancer Models
• Chemistry
• Ketcher Chemistry
• Human Genetics
• Genomes
• RNA
• CellGuide
• ZINC

The important point is not just the length of the list. It is the range of real research tasks those connectors cover:

Runcell Science also includes session-level connector management. A session connectors menu can enable or disable bundled science connectors and configured MCP servers for the current session. That makes the connector layer task-specific. Instead of handing every possible tool to the agent, you can expose the sources that fit the research question in front of you.

Runcell Science vs Claude Science

People searching for "open source alternative to Claude Science" usually care about three questions:

1. Can the tool support ongoing scientific work instead of one-off answers?
2. Can it help write, inspect, and maintain research code?
3. Can the team avoid being locked into one vendor, one model, or one hosted workspace?

Runcell Science answers those questions with an open workspace model.

There is an important caveat: Runcell Science is still early and moving quickly. It is better suited to technical users who are comfortable trying a local development environment, working with agent CLIs, and evaluating open source tools. It is not yet a fully packaged enterprise SaaS that covers every scientific use case out of the box.

Example Research Workflows

Literature review: from search to reusable notes

In a literature-heavy session, a researcher could enable PubMed, bioRxiv / medRxiv, and Literature Graph connectors. The agent can help search a topic, extract key papers, compare research directions, and turn the results into a reusable document or artifact.

The goal is not to ask a model to summarize papers from memory. The goal is to let the agent work around actual retrieval results, then keep the summary, notes, and next steps inside the project workflow.

Target and compound screening

Drug discovery and bioinformatics teams may want ChEMBL, Genes & Ontologies, Protein Annotation, Structures & Interactions, Variants, ZINC, and Chemistry connectors in the same session.

A typical workflow might look like this:

1. Query genes and ontology information for a target.
2. Inspect protein annotation and structural interaction context.
3. Retrieve known compounds and activity data.
4. Generate an analysis notebook or prototype script.
5. Review the diff and capture follow-up tasks.

This is where an AI coding agent becomes more useful than a chat assistant. It can move from retrieval to code to artifacts while staying attached to the research project.

Clinical trial research inside a project context

For clinical and translational questions, Clinical Trials and Drug Regulatory connectors can help the agent work around structured trial and regulatory information. Runcell Science's advantage is that those findings can be connected to local notebooks, scripts, reports, and artifacts instead of ending as a pasted chat summary.

Variant interpretation and analysis outputs

Variants, Human Genetics, Genomes, Regulation, and Expression connectors can support exploratory work around genetic variants, expression context, and regulatory background. The agent can turn those queries into a notebook draft, markdown report, visualization outline, or validation checklist.

Notebooks, pipelines, and reproducible research artifacts

Runcell Science is not only about notebooks, and it is not only about code generation. Its workspace model fits the messy middle of research work: scripts, notebooks, docs, figures, experiment notes, and project state.

It can help an agent:

• draft analysis code and notebook cells;
• organize experiment steps and README files;
• explain pipeline errors;
• generate prototypes;
• review code diffs;
• turn intermediate results into reproducible artifacts.

If your work happens primarily inside a live Jupyter runtime, read Jupyter AI with RunCell. If your focus is a research repository with connectors, agent sessions, diffs, and project-level workflow, Runcell Science is closer to the AI research workspace pattern.

Current Capabilities vs Natural Extensions

It is useful to separate what Runcell Science clearly supports today from where the architecture could expand.

This distinction keeps the positioning honest. Runcell Science should not be read as a claim that one open source project already replaces every scientific software tool. Its promise is more specific: a local-first, connector-first workspace where AI coding agents can participate in real research projects.

Who Should Try Runcell Science Now?

Runcell Science is a strong fit if you:

• want to use Codex or Claude Code for research code without scattering work across terminals and chats;
• need to connect papers, data sources, notebooks, pipelines, and artifacts;
• are exploring scientific AI agents or AI research workspaces;
• want to build custom MCP connectors or private research tool layers;
• prefer local project control over a fully hosted research context.

It may not be the right choice yet if you:

• need a zero-configuration commercial scientific AI product;
• only want a quick literature summary;
• do not want to install or authenticate local agent CLIs;
• need mature enterprise permissions, audit controls, and SLAs before trying the workflow.

How to Start Locally

Runcell Science is open source on GitHub:

https://github.com/runcell-ai/runcell-science (opens in a new tab)

The local start command from the project is:

./scripts/dev.sh

Then open the local web app:

http://127.0.0.1:27183

For agent-backed sessions, you also need the relevant local agent CLI installed and authenticated, such as codex or claude. That fits the product model: the workspace is local-first, and the agent backend is selected according to your environment and preferences.

Related Guides

• 2026 Best AI Coding Tools
• Codex vs Claude Code Skills
• Can Claude Code Analyze Jupyter Notebooks for Data Science?
• Jupyter AI with RunCell
• How to Build an MCP Server for Your Blog Website
• Agentic RAG

FAQ

Is Runcell Science an open source alternative to Claude Science?

Yes, Runcell Science fits the search intent for an open source alternative to Claude Science, but the product model is different. It emphasizes a local-first workspace, replaceable agent backends, session-level connector management, and reviewable research artifacts rather than a single locked agent experience.

Can Runcell Science use Codex or Claude Code?

Yes. The current runtime supports codex and claude agent providers, which correspond to Codex and Claude Code workflows. The workspace is designed so the research environment is not tied to only one agent backend.

Could Runcell Science support Gemini or other code agents later?

The architecture points in that direction because the agent backend is replaceable. Today, the concrete providers are Codex and Claude Code. Future support for Gemini, private agents, or other code agents would depend on runtime integration, permissions, tool protocols, and product implementation.

What research connectors does Runcell Science include?

The bundled science connectors include PubMed, bioRxiv / medRxiv, Clinical Trials, ChEMBL, BioMart, Genes & Ontologies, Protein Annotation, Structures & Interactions, Variants, Expression, Omics Archives, Drug Regulatory, Chemistry, Ketcher Chemistry, ZINC, and related research resources. It can also work with configured MCP servers.

How is Runcell Science related to open science?

Runcell Science is not the same thing as open science itself. It belongs to the open science tooling category because it supports local control, reproducible artifacts, auditable agent sessions, open connector workflows, and reviewable code diffs.

Why do session-level connectors matter for research agents?

Scientific agents need access to real sources, not just model memory. Session-level connector management lets a researcher expose the right literature, clinical, genomics, chemistry, or MCP tools for a specific task while keeping unrelated tools out of the session.

Is Runcell Science production-ready?

It is best viewed as an early open source project for evaluation, local pilots, and technical exploration. Teams with strict enterprise requirements should test it on a narrow research workflow before relying on it broadly.

Bottom Line

The next useful step for scientific AI is not another general chatbot. The harder problem is connecting project context, research data sources, agent execution, generated artifacts, code diffs, and reproducible workflow.

Runcell Science is interesting because it combines those pieces in a local-first, connector-first workspace. If you are evaluating Claude Science and want an open source alternative, start with the Runcell Science GitHub repository (opens in a new tab), run it locally, and test it against a real research repo. The important question is not whether the demo looks impressive. It is whether the agent can understand your project, use the right connectors, produce reviewable diffs, and keep the next research step inside the same session.