Set up ToolUniverse for Claude Science

The upstream ToolUniverse ships a Claude Code plugin (MCP server + uvx + slash commands). Claude Science loads capabilities differently, so this skill installs the equivalent natively: the tooluniverse pip package supplies the 2500+ tools, and the workflow library is packaged into a single dynamically-loaded skill, tooluniverse-research. No uv, no MCP server, no plugin marketplace.

Loading this skill defines tu_build_research_bundle() in the kernel (run cells in the tooluniverse conda env).

Full install / update — four steps

1. Create the conda env (skip if it already exists):

manage_environments(mode="create", name="tooluniverse", python_version="3.11", packages=["pip"])

2. Install (or upgrade) the tools — the pip package is the tool layer:

manage_packages(mode="install", environment="tooluniverse", packages=["tooluniverse"], use_pip=True)

Pin a version for reproducibility with ["tooluniverse==1.3.0"].

3. Stage the workflow bundle — fetch the current repo and rebuild the file tree (run in a python cell, env tooluniverse):

res = tu_build_research_bundle(staging="./tu_staging")
res  # {out_dir, n_workflows, n_files, dropped, files_head}

This downloads the repo tarball, parses every tooluniverse-* workflow (dropping the plugin/installer entries), and writes ./tu_staging/out/ = SKILL.md, kernel.py, index.json, workflows/*.md.

4. Publish the skill — push the staged tree into the catalog (run in the repl tool; host.skills.* lives there, not in python):

import os
SKILL = "tooluniverse-research"
out = os.path.abspath("./tu_staging/out")
if any(s["name"] == SKILL for s in host.skills.list()):
    host.skills.delete(SKILL)                       # clean rebuild
for root, _d, fs in os.walk(out):
    for f in fs:
        p = os.path.join(root, f)
        rel = os.path.relpath(p, out)
        host.skills.edit(SKILL, rel, open(p, encoding="utf-8").read())
print(host.skills.publish(SKILL, overwrite=True))

(host.skills.publish refuses if kernel.py fails the sidecar gate — the edit result carries the verdict.)

Verify

skill("tooluniverse-research")                      # loads router + injects helpers
tu = get_tu()
tu.run({"name": "PubChem_get_CID_by_compound_name", "arguments": {"name": "metformin"}})
# -> {'status': 'success', 'data': {'IdentifierList': {'CID': [4091]}}}

Notes

• Sandbox cache: ToolUniverse defaults its cache to ~/.tooluniverse, which is read-only here; get_tu() redirects it to the workspace via TOOLUNIVERSE_CACHE_DIR. Nothing to configure.
• API keys (optional): most tools work without them. For NCBI / OncoKB / NVIDIA etc., add keys under Customize → Credentials, then expose them in the tooluniverse env.
• What is NOT ported: the plugin's slash commands (/tooluniverse:research) and MCP server — replaced by natural-language routing (search_skills → find_tu_workflow). The two *-plugin installer docs are dropped as non-research entries.
• Updating: rerun steps 2–4. Step 2 upgrades the tools; steps 3–4 refresh the workflow library from the latest GitHub state.