Skip to content

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 Review Every Action 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.

Request Focused Research Review

The parent agent can delegate one bounded task at a time to advisory specialists for research planning, data quality, cohort and feature definitions, statistical methods, or reproducibility. These research specialists use the active model and return their review to the same conversation without receiving project tools.

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 the Stanford AI API Gateway, sign in with an eligible ChatGPT account, provide an OpenAI or Anthropic API key, connect another authorized OpenAI-compatible service, or select 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