host-and-share-remote-tool

Host and share a remote tool

Use this workflow for user-owned code and infrastructure. Keep the workload on the user's computer or lab server; ToolUniverse Platform receives only the MCP tool manifest and relayed calls.

Choose the shortest path

• Python function, model, database query, or workflow: wrap only the callable with @remote_tool, then use tu serve ... --share.
• Existing Streamable HTTP MCP server: keep it running on loopback and use tuplatform-relay --forward ....
• One of ToolUniverse's 30 reviewed scientific implementations: use its implementation-specific setup-<name>-remote-tool skill and tu remote share <name> instead.

Do not treat an arbitrary REST endpoint as MCP. Wrap it in a typed Python function first, or put a reviewed MCP adapter in front of it.

1. Install the reviewed clients

Open ToolUniverse Connect, go to Computers → Connect a computer (https://connect.aiscientist.tools/remote-servers), choose the Python or existing-MCP path, and copy the immutable install command shown there. Run it in a new Python 3.12 virtual environment. The command pins both the ToolUniverse and relay sources; do not replace the pins with a moving branch.

Confirm the expected commands exist:

tu --help
tuplatform-relay --help

Stop if installation or source access fails. A locally working model does not prove that sharing works.

2A. Wrap a Python model or function

Create my_tool.py. Load fixed model artifacts from provider-owned configuration; do not accept arbitrary caller-controlled filesystem paths or model identifiers.

from tooluniverse import remote_tool

@remote_tool
def predict(sequence: str, threshold: float = 0.5) -> dict:
    """Score one sequence with the locally hosted model."""
    if not isinstance(sequence, str) or not 1 <= len(sequence) <= 10_000:
        raise ValueError("sequence must contain 1 to 10,000 characters")
    score = min(len(sequence) / 100.0, 1.0)  # replace with the real model call
    return {"score": score, "passes": score >= threshold}