The MCP server is launched with uvx, so uv must be installed. If it isn't, the tooluniverse MCP server cannot start and its tools won't appear (the plugin prints a one-line reminder on startup).

curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS/Linux
# Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Then restart Claude Code. (No separate tooluniverse install needed — uvx fetches it on first use.)

Install

# 1. Register the marketplace from GitHub
claude plugin marketplace add mims-harvard/ToolUniverse

# 2. Install the plugin
claude plugin install tooluniverse@tooluniverse

Restart Claude Code. Done.

Full install guide (troubleshooting, API keys, offline zip install, version pinning): see the tooluniverse-claude-code-plugin skill.

Local development install

# From the repo root:
claude plugin marketplace add ./
claude plugin install tooluniverse@tooluniverse

What's Included

MCP Server: Auto-configured via .mcp.json. Provides find_tools, list_tools, get_tool_info, execute_tool — accessing 1000+ scientific APIs.
120+ Research Skills: Specialized workflows for genomics, drug discovery, clinical analysis, statistical modeling, data wrangling, and more.
Slash Commands (each enforces a discipline the agent won't apply on its own):
- /tooluniverse:research — drive a multi-database investigation inline in this chat (you see every step)
- /tooluniverse:translate-id — resolve an ID across all relevant namespaces
- /tooluniverse:cross-validate — verify a claim across 3+ independent databases
- /tooluniverse:compare — N-way side-by-side comparison with domain-appropriate columns
- /tooluniverse:literature-sweep — graded mini-review across PubMed + EuropePMC + Semantic Scholar
- /tooluniverse:self-review — review current or supplied work against its actual goal; qualitative by default, scoring only when requested
- /tooluniverse:setup-keys — configure ToolUniverse API keys
Research Agent: /tooluniverse:researcher — same investigation, delegated to a forked-context subagent that returns one summary (use when you don't want intermediate tool-call output in your main thread).

Usage

Default: just ask. The router skill auto-dispatches scientific questions to the matching sub-skill — no command prefix needed.

What do we know about drug resistance mutations in EGFR?
Compute differential expression for these RNA-seq counts.
What's the prevalence of distal renal tubular acidosis?

For specific surfaces, use the right interaction below.

Discover and run tools (no slash command needed)

Use this	When
Natural language ("what's the prevalence of X?")	Default. The router skill auto-dispatches; the agent finds and runs tools itself.
`find_tools("topic")` (MCP, agent-side)	One-shot keyword search, you already know the topic
`tu find "topic"` (CLI)	Same, from the shell
`tu run <name> '<json>'` (CLI)	You know the exact tool name and JSON args. Fastest. Good in scripts.

Slash commands (when discipline matters)

Use this	When
`/tooluniverse:research <question>`	You want to drive a multi-database investigation inline in this conversation, watching each step. Enforces the same look-up-don't-guess + cross-validate + honest-INDETERMINATE discipline as the `researcher` agent, but stays in-thread so you can interrupt and refine.
`/tooluniverse:translate-id <id>`	You have an ID in one namespace and need it in others (HGNC ↔ Ensembl ↔ UniProt ↔ NCBI ↔ OMIM ↔ ChEMBL ↔ PubChem). Detects the input namespace, picks the right resolver, returns a complete cross-reference table.
`/tooluniverse:cross-validate <claim>`	You have a specific testable claim and want to know how strongly it's supported. Forces 3+ independent databases, reports concordance per source. Use before publishing or acting on a fact.
`/tooluniverse:compare <items>`	You want a side-by-side comparison of N items (drugs, targets, diseases, variants). Detects the domain, picks domain-appropriate columns, produces a tabular report.
`/tooluniverse:literature-sweep <topic>`	Graded mini-review across PubMed + EuropePMC + Semantic Scholar with dedup, relevance scoring, and a recommended reading order.
`/tooluniverse:self-review [target or request]`	Review current or supplied work against the original goal and surface evidence-backed findings, strengths, gaps, and fixes. Plain eval/review requests stay qualitative; ask explicitly for a score, grade, weighted rubric, Qworld, or RET when numeric grading is wanted. Omit the argument to review the current work from conversation context.
`/tooluniverse:setup-keys`	Set up ToolUniverse API keys — opens a setup page listing every key tools can use, with registration links, and saves them to a `.env` the MCP server, CLI, and SDK all read.

Launch the research agent

Use this	When
`/tooluniverse:researcher <question>`	Same investigation as `research`, but delegated to a forked-context subagent that runs in isolation and returns one summary. Use when the research will otherwise pollute your main conversation with many intermediate tool calls, or when you want a clean self-contained answer (e.g., to paste elsewhere). For in-thread step-by-step research with follow-ups, use `research` instead.

Bulk tool calls (5+ in a loop)

from tooluniverse import ToolUniverse
tu = ToolUniverse()
tu.load_tools()
for gene in ["TP53", "BRCA1", "PIK3CA", "PTEN", "KRAS"]:
    print(tu.run_one_function({"name": "ensembl_lookup_gene",
                                "arguments": {"gene_id": gene, "species": "homo_sapiens"}}))

Use the SDK over tu run in a Bash loop — registry loads once, avoiding per-call startup overhead.

Design Philosophy

The router skill is the front door. Just ask scientific questions in natural language; the router auto-routes to the specialized skill. No slash command needed for normal research.
tu run is the fastest tool surface when you know what you want.
The slash commands earn their weight only when they enforce a discipline the agent won't apply on its own — multi-source verification (cross-validate), namespace-complete ID resolution (translate-id), structured N-way comparison (compare), graded literature triage (literature-sweep). Wrappers around defaults don't earn a command.
MCP (find_tools, get_tool_info, execute_tool) is what the agent uses internally — exposed for direct use too, but tu run from Bash is usually simpler.

API Keys (Optional)

Most tools work without API keys. For enhanced access:

Key	Source	Free?
`NCBI_API_KEY`	https://www.ncbi.nlm.nih.gov/account/settings/	Yes
`SEMANTIC_SCHOLAR_API_KEY`	https://www.semanticscholar.org/product/api	Yes
`ONCOKB_API_TOKEN`	https://www.oncokb.org/apiAccess	Academic

Set via environment variables or in the MCP server config.

Updating

Recommended: enable auto-update so new versions install on Claude Code restart.

Run /plugin in Claude Code
Go to the Marketplaces tab
Find tooluniverse and toggle Enable auto-update ON

Once enabled, Claude Code re-pulls this repo's marketplace.json on each startup, compares the plugin's version field, and silently downloads any newer release. You'll get a notification prompting /reload-plugins after the new version is staged.

To update manually:

/plugin marketplace update tooluniverse
/plugin uninstall tooluniverse@tooluniverse
/plugin install tooluniverse@tooluniverse
/reload-plugins

What updates together with the plugin — skills, slash commands, the research agent, hooks, and the MCP server. There's no separate update mechanism for any of them; they share the plugin's version lineage.

The MCP server's Python package (tooluniverse on PyPI) is invoked via uvx tooluniverse in the bundled .mcp.json. uvx caches the resolved package and refreshes it periodically per its own cache-eviction policy — you get new PyPI releases without paying a re-resolve cost on every session start, and offline sessions still work from the cache.

To force a refresh now, run uvx --refresh tooluniverse --version from a shell once; the cache update then applies to future MCP launches.

To pin a specific version, edit your .mcp.json to "args": ["tooluniverse@1.2.0"] (or any released version). Useful for reproducibility on long-running analyses.

To opt out of auto-updates: leave the marketplace's auto-update toggle OFF, or set DISABLE_AUTOUPDATER=1 + FORCE_AUTOUPDATE_PLUGINS=1 env vars to disable Claude Code auto-update while keeping plugin auto-update enabled (or vice versa).

README.md

ToolUniverse Plugin for Claude Code

Prerequisite: uv