Heartwood¶
Heartwood is an open-source coding agent for biomedical research projects. Describe the work you need in ordinary language, inspect the files and commands the agent proposes, and keep a reviewable history of the session inside your research environment.
Heartwood is designed for workstations, containers, Terra, Stanford Carina, and operator-managed research platforms. It can use models provided by a research environment, hosted model services, another compatible service, or a model managed by Heartwood in the same compute environment.
Start Your First Project Choose a Platform
What You Can Do¶
Work Through a Research Task¶
Ask Heartwood to inspect code, prepare an analysis, explain a result, or update project files. Heartwood treats the folder from which you start it as the project, so begin in a dedicated folder that contains only the material intended for the session.
Review Actions Before They Run¶
Heartwood presents related proposed actions together before execution under the default Ask Every Time setting. You can allow or reject the complete set after inspecting every proposed command and file operation.
Reuse Research Skills¶
Repository-verified Skills provide reusable instructions and tools for supported research workflows. Projects can add explicitly reviewed Skills for additional workflows.
Continue Across Interfaces¶
The terminal, browser, and notebook bridge use the same project configuration, model selection, sessions, action decisions, and audit history when the platform supports those interfaces.
Start With Your Environment¶
| Where You Work | Recommended Starting Point | Available Interfaces |
|---|---|---|
| macOS or Linux workstation | Heartwood container | Terminal and browser; Jupyter not included |
| Existing Python development environment | Native development install | Terminal, browser, notebook |
| Terra workspace | Heartwood for Terra | Terminal and notebook |
| Stanford Carina | Heartwood for Carina | Terminal |
| Managed research platform | Deployment guidance | Defined by the platform deployment |
Choose Where the Model Runs¶
Heartwood does not include model weights or credentials in its images or installers. During setup, choose an available research-environment connection, OpenAI, Anthropic, another authorized OpenAI-compatible service, or a model that Heartwood manages in the current environment.
The platform remains authoritative for identity, network access, data permissions, and which model routes may receive project content.
Understand the Boundary¶
Heartwood can operate inside a reviewed environment, but installing it does not authorize access to protected health information, approve a model provider, or make a deployment compliant. Your institution remains responsible for the research platform, agreements, access controls, storage, networking, and data handling rules.
Begin with synthetic or non-sensitive files until the complete deployment has been reviewed for the intended data. See Security and Controlled Data before using restricted data.
Find an Answer¶
- Your First Project provides the shortest complete path from installation to a reviewed action.
- Work With Heartwood explains the normal conversation, action, replay, and audit workflow.
- Models compares research-environment, hosted, compatible-service, and Heartwood-managed routes.
- Platforms covers containers, Terra, and Stanford Carina.
- Diagnostics and Troubleshooting maps stable
HW-*codes to recovery steps. - How Heartwood Works explains the architecture and security boundaries.