Claude Science on Chimera

Access: JupyterHUB

Claude Science is Claude's agentic research application. On Chimera it runs as a proxied web application inside your JupyterHUB session, using a shared deployment under /work/_software/claudescience. You do not install the application itself, only a small launcher package that adds the tiles to your JupyterLab launcher.

You sign in with your own Claude account and subscription. Chimera does not provide a shared Claude subscription or a shared API key.

The launcher package provides two tiles, and they are different products in practice:

Tile Chimera MCP tools Sandbox Documented on
claude-science Enabled No this page
claude-science-SANDBOX Disabled Yes Claude Science SANDBOX

This page covers the Chimera-integrated claude-science tile, and the installation that both tiles share. For a comparison with the other platforms, see Agentic platforms for research.

Prerequisites

  • A Chimera account with access to JupyterHUB.
  • A writable /work/$USER directory. Claude Science keeps all of its state there and never falls back to /home, so it cannot start without it. Check with test -w "/work/$USER" && echo OK. If it is missing or not writable, ask the Chimera administrators to create it for your account before going any further — the installer refuses to run without it, and the tile would fail when you clicked it.
  • Your own Claude account with an active subscription.

Install the launcher

Claude Science is not installed for you. Do this once. One installation provides both tiles.

  1. In JupyterLab, click the Terminal app.
  2. Type:
/work/_software/claudescience/install-launcher

The script installs the newest version. Run it again later to update.

This per-user installation is the current arrangement. A shared, global installation is being considered, which would remove this step.

Start Claude Science

  1. Restart your JupyterHub server, so that JupyterLab discovers the new launcher entries. Reloading the browser page is not enough.
  2. Open JupyterLab again.
  3. In the launcher, click the claude-science tile.

The first start can take a while — the application warms up its tooling before the interface appears, and the launcher allows up to 900 seconds for it. If something is actually wrong, you get an error page naming the problem instead of a wait.

First sign-in

On first use, Claude Science opens a sign-in dialog. Follow the link to Claude.ai, authorize the application, then paste the authorization code back into Claude Science.

Each tile signs in separately, because each keeps its own profile directory.

The claude-science tile

This is the Chimera-integrated launcher. Use it when you want Claude Science to work with Chimera itself. It has these integrations enabled:

  • Chimera Slurm MCP
  • Chimera FileCompress MCP
  • Chimera-related skills

Through these, you can ask Claude Science to inspect the queue, submit and diagnose jobs, read job logs, work with files in your work area, and handle archives. What those tools can and cannot do is documented in full on the AI Interfaces for Chimera page.

The launcher loads the MCP servers through the cluster module system, roughly:

module load chimera-mcp
chimera-slurm-mcp
chimera-filecompress-mcp

It initializes the module system explicitly, because Claude Science starts MCP servers from a non-interactive shell where the module function is not defined automatically.

The launcher merges the managed chimera-slurm and chimera-filecompress entries into your profile's MCP configuration. Any other MCP servers you have configured yourself are preserved, unless they use those same managed names.

Warning

This tile is not sandboxed. It runs without the Claude Science sandbox, because the Chimera MCP tools need to start external command-line processes through the cluster module system.

This means Claude Science runs with your own Unix permissions and can reach any file your account can reach. Treat it as an agent acting as you. It can also submit and cancel jobs and consume your allocation, and anything it reads may be sent to Claude, including file paths, log contents, job metadata, and parts of your data.

If you want isolation and do not need the Chimera tools, use the Claude Science SANDBOX tile instead.

Data location

The claude-science tile keeps its state in:

/work/<username>/.claude-science

It stores login and session state, your selected Claude organization and subscription state, local runtime files, logs, MCP configuration, caches, and Claude Science project data there.

The launcher deliberately refuses to use /home for this state, and redirects the runtime environment into the profile directory to keep the smaller /home filesystem from filling up:

HOME
TMPDIR
XDG_CONFIG_HOME
XDG_CACHE_HOME
XDG_DATA_HOME

The shared deployment under /work/_software/claudescience holds the shared command, deployment scripts, the launcher wheel, default MCP configuration, and shared skill sources. Do not store your own data there; your state belongs under /work/<username>.

Troubleshooting

The tile does not appear in the launcher. Restart your JupyterHub server; a browser reload does not pick up new launcher entries. If the tiles are still missing, check what is installed:

/usr/bin/python3 -c 'import importlib.metadata as m; print(m.version("jupyter-claudescience-proxy"))'

The tile shows an error page instead of starting. The page names the cause. By far the most common one is a missing or unwritable /work/<CASlogin>; see Prerequisites. Once the cause is fixed, restart your JupyterHub server — reloading the page is not enough, because the launcher process starts only once per session.

First sign-in returns to the sign-in screen. Claude Science restarts its own daemon once your organization is selected, which happens right after the first sign-in. Launcher versions before 0.1.13 mistook that restart for a crash, which cost a slow recovery in the middle of signing in. Upgrade to 0.1.13 or newer. If it still happens, wait a moment and refresh the same URL.

The Chimera MCP tools are missing. Check that the chimera-mcp module can be found on the node where your JupyterHub server is running:

module spider chimera-mcp

Note that the sandboxed tile does not load the Chimera MCP tools at all; that is intentional.

You need the details behind a failure. Everything the launcher and your Jupyter server print goes to one file in your home directory, ~/jupyterhub_slurmspawner_<jobid>.log. Open the newest one with ls -t ~/jupyterhub_slurmspawner_*.log | head -1. Nothing in JupyterLab surfaces this file, but it is the first place to look for any startup problem.

The session keeps running after you close the browser. It does. Closing the tab or using Log Out does not stop your JupyterHub session, and it continues to hold cluster resources. End the session from the JupyterHub control panel, as described in JupyterHUB.