OpenScience can read a project, find relevant files, edit code, and run commands with the tools and access available to the session. Use it to understand an unfamiliar analysis repository or improve the code supporting your research.
Map an unfamiliar repository
Read the README, project instructions, and dependency files. Map the main
entrypoints, data inputs, analysis steps, outputs, and tests. Save a short
project guide with exact paths. Do not install dependencies or change
code until we agree on the next task.
A local code map comes from reading and searching the project. It does not imply that a separate hosted repository index has been created.
Ask for a focused change
Provide the observed behavior, expected behavior, and reproduction command:
The analysis command fails when a sample has a missing batch label.
Reproduce the issue with the existing fixture, fix the behavior, and run
the relevant tests. Explain how missing labels are handled in the output.
Ask for a diff and validation results before using the updated analysis. Keep changes to unrelated files out of a focused repair.
Use the project terminal
Open Terminal in the project or use Open project terminal from the command palette. You can run commands yourself while following the research conversation.
New terminals open in the selected Working folder, or in session scratch when you explicitly select scratch. With no connected folder, they open in the project directory. Existing terminals keep their current directory when you change the selection; open a new tab to use the new folder.
This is your normal local shell, with your login configuration, installed commands, credentials, and operating-system permissions. Commands such as claude, git, and your package managers can use their usual files and network connections. On a remote OpenScience server, the terminal runs on that server. The research agent's shell commands and Python/R kernels continue to use the agent's sandbox and approval settings.
Use Control+backtick to open or close the terminal and Control+Shift+backtick for a new terminal when available. See Keyboard shortcuts.
Confirm the working directory and selected environment before running a command. The terminal's environment need not contain every package available to the agent's Python or R analysis environment.
Use Claude Code with its own login
Choose Claude Code beside New in the Terminal panel to launch your installed claude command in the selected working folder. This opens the interactive CLI: Claude handles its own login, streaming output, tool approvals, and conversation history. It uses your Claude account, without requiring an Anthropic API key in OpenScience.
Install Claude Code on the machine running the workspace first. If the shell reports claude: command not found, install it there or fix that shell's PATH, then run claude in the same tab. Follow the CLI's sign-in instructions if you are not logged in.
Use Claude's interruption controls to stop a response and /exit to return to the shell. Run claude --continue to continue the most recent conversation in that folder, or claude --resume to choose a conversation. Closing the terminal stops its processes; reconnecting starts a new CLI process and does not automatically replay a prompt. The research chat's model selection is independent of this terminal session.
Claude Code launched here has the same normal local-terminal access as a command you type yourself. OpenScience's research-agent tools continue to use their configured sandbox. On a remote workspace, install and sign in to Claude on that server.
Keep version control deliberate
Ask OpenScience to inspect the working tree before modifying an existing repository. Describe whether you want a local change, commit, branch, or pull request, and which actions require your review.
Implement the fix on a new branch. Preserve existing uncommitted work,
run the relevant tests, and show the diff. Do not push until I ask.
Project instructions in AGENTS.md can record routine checks and conventions. A branch or commit is useful for code history; it is not a backup of external data or services.
Reproduce the final state
The report should include changed behavior, files, commands run, test results, and limitations. For a research script, also record data and dependency versions. Use Automation when the same bounded task should run through a script or CI job.