Tools and environments changedclaude-science/tools-and-environments
Nearest release: v2.1.302, published under an hour before upstream edited the page. Shown because the two are within 24 hours of each other. Nothing here says the release caused the edit.
Upstream edited this page at 16 Oct 2026 17:16 UTC, give or take a minute or two: the time comes from Anthropic’s own sitemap rather than from a commit. This site recorded the change at 16 Oct 2026 17:37 UTC.
Upstream edited
Recorded here
Lines+38added
Lines−3removed
From line
31
where the diff opens
First seen
14 Aug 2026
this site's first read of the page
Recorded edits5to this page, all time
### Environments in granted folders ### How Claude Science detects a GPU
The whole hunk
from line 31, old and new numbered
/
from line 31
3131
3232The sandbox has no root access or system package manager. `apt` and `sudo` aren't available; Claude uses conda-forge or builds from source instead.
3333
34### Environments in granted folders
35
36On Linux, a uv, conda, mamba, or pixi environment in a folder you've granted **Read & write** can still be read-only to Claude. This happens when the environment's package files are hard links shared with the tool's package cache, the folder where the tool keeps the packages it has downloaded. A change to those files could also change the cache and other environments that share them.
37
38Claude can still run code from the environment but can't add, update, or remove packages in it. The failure shows as `Read-only file system` from pip, uv, and pixi, and as `EnvironmentNotWritableError` from conda. In a uv project, `uv run` and `uv sync` fail the same way whenever they have to update the environment, for example after `pyproject.toml` changes.
39
40To add packages, choose one of these:
41
42* **In a uv project, ask Claude to use a separate environment.** Have Claude set `UV_PROJECT_ENVIRONMENT` to the full path of a folder in the session's [workspace folder](/docs/claude-science/core-concepts#projects-and-sessions) and run uv as usual. The variable has to stay set for later uv commands, including `uv run` and `uv sync`. The separate environment belongs to that session, so another session starts without it. For a change every session keeps, change the environment yourself in a terminal outside Claude Science.
43* **With any of these tools, change the environment yourself.** In a terminal outside Claude Science, use the tool's own commands to add, update, or remove packages.
44
3445## GPUs
3546
36If your Linux machine has GPUs, the sandbox makes them available to code Claude runs, including on multi-GPU machines. You first need to turn GPU access on in the Settings > Compute pane.
47If your Linux machine has NVIDIA GPUs, the sandbox makes them available to code Claude runs, including on multi-GPU machines. To use them, turn on the **Host GPU** switch in **Settings > Compute**. Sessions you start after that can use the GPU. A session started while the switch was off keeps running without it, and so does a session continued from it with **Continue as a new session**.
3748
3849<Note>
39 Allowing GPU access reduces the default sandboxing configuration applied by Claude Science.
50 Turning on the **Host GPU** switch reduces the default sandboxing configuration applied by Claude Science.
4051</Note>
4152
42If your machine has no GPU, Claude notes this when a task needs one and can run the work on the remote compute you've connected. See Remote compute clusters and External compute providers.
53If Claude Science finds no GPU, Claude notes this when a task needs one and can run the work on the remote compute you've connected. See [Remote compute clusters](/docs/claude-science/remote-compute-clusters) and [Compute providers](/docs/claude-science/compute-providers). If your machine has an NVIDIA GPU but the **Host GPU** switch isn't in **Settings > Compute**, or Claude says there's no GPU in a session started with the switch on, see [How Claude Science detects a GPU](#how-claude-science-detects-a-gpu).
54
55### How Claude Science detects a GPU
56
57On Linux, Claude Science detects an NVIDIA GPU by running `nvidia-smi`. When it finds no GPU this way, Claude treats the machine as having no GPU, and the **Host GPU** switch doesn't appear in **Settings > Compute** unless it's already on.
58
59Claude Science runs the first `nvidia-smi` it finds in these folders, in this order:
60
611. `/usr/local/nvidia/bin`
622. `/usr/bin`
633. `/bin`
644. `/usr/lib/wsl/lib`
655. `/opt/bin`
66
67To see which folder yours is in, run `command -v nvidia-smi` in a terminal outside Claude Science. Claude Science doesn't look for `nvidia-smi` in any other folder, such as `/usr/local/bin`. If the `nvidia-smi` in a listed folder is a link, the file it finally leads to, after any other links, must also be named `nvidia-smi` and be under `/usr`, `/lib`, `/lib64`, `/bin`, `/sbin`, or `/opt`. To see the file a link finally leads to, run `readlink -f` with its path, for example `readlink -f /usr/bin/nvidia-smi`.
68
69`nvidia-smi` also has to work without the `LD_LIBRARY_PATH` environment variable, because Claude Science doesn't pass that variable on to it, even when your terminal has it set. To check, run it in a terminal outside Claude Science by its full path with the variable removed, for example `env -u LD_LIBRARY_PATH /usr/bin/nvidia-smi`. If that fails where plain `nvidia-smi` works, the variable is the reason. Make `nvidia-smi` work without the variable, or ask your system administrator to.
70
71If `nvidia-smi` is in another folder, link it into `/usr/bin`, which takes administrator (`sudo`) access. If you don't have that access, ask your system administrator to make the link. For example, when `nvidia-smi` is in `/usr/local/bin`, run this in a terminal outside Claude Science:
72
73```bash theme={null}
74sudo ln -s /usr/local/bin/nvidia-smi /usr/bin/nvidia-smi
75```
76
77After either fix is in place, run `claude-science stop` in a terminal outside Claude Science and start Claude Science again. If Claude Science then finds the GPU, the **Host GPU** switch is in **Settings > Compute** and its label shows the GPU's name. Turn it on if it's off, then start a new session, not one made with **Continue as a new session**.
4378
No line in this hunk matches that.