Launch and Share a Tool

ToolUniverse can turn a typed Python function into a local Streamable HTTP MCP server. The same command can connect that server to ToolUniverse Platform without opening a public inbound port on the provider machine.

Define a Tool

Create my_model.py:

from tooluniverse import remote_tool

@remote_tool
def predict(sequence: str, threshold: float = 0.5) -> dict:
    """Run my sequence model."""
    return model.predict(sequence, threshold=threshold)

@remote_tool derives the MCP input schema from type hints and defaults. It returns the original function unchanged, so ordinary unit tests can call predict(...) directly.

Run Locally or on a LAN

tu serve my_model.py

The default endpoint is http://127.0.0.1:8080/mcp. To make it reachable on a trusted LAN, opt in to a non-loopback bind:

tu serve my_model.py --host 0.0.0.0 --port 8080

Do not expose an unauthenticated development server directly to the public internet.

Connect the Tool from Another Machine

Register the MCP endpoint once:

tu connect http://gpu-host:8080/mcp --name my-model

On later ToolUniverse.load_tools() calls, the remote operation is discovered and appears with the stable prefix my_model_.

Each connection name must generate a unique prefix; ToolUniverse rejects a second connection that could silently replace tools from the first. Inspect or remove saved connections without contacting their servers:

tu connections
tu disconnect http://gpu-host:8080/mcp

For a bearer-authenticated MCP endpoint, store only the environment-variable name, never the secret itself:

export MY_MCP_TOKEN=...
tu connect https://gpu.example/mcp --name my-model --auth-env MY_MCP_TOKEN

Share Without a Public IP

For a provider-only environment, use Python 3.11 or 3.12 and install the small runtime instead of ToolUniverse’s complete scientific dependency collection:

python -m pip install fastmcp pyyaml "tuplatform-connect @ git+https://github.com/tooluniverse/tuplatform.git@fde2edb73d573fac3fe8831e9ca19771b7583931#subdirectory=sdk/tuplatform-connect"
python -m pip install --no-deps tooluniverse

Until this feature reaches a PyPI release, use the reviewed ToolUniverse source command shown in ToolUniverse Connect instead of the second command above. If the full ToolUniverse SDK is already installed, only the relay package is needed. This reduced installation is intended for running provider files with the tu serve command; install the normal ToolUniverse dependencies when the provider code also uses the full Python SDK.

Then run:

tu serve my_model.py --share --name "My sequence model"

Validate both the local MCP manifest and the computer-only platform connection before sharing:

tu doctor --forward http://127.0.0.1:8080/mcp

Enter a computer-only connection key at the hidden prompt. For an unattended service, provide it in a protected TOOLUNIVERSE_SERVICE_KEY environment variable. The key is not written into the connection file or the command line.

The provider machine makes the outbound WebSocket connection; no inbound firewall rule or public IP is required. Open My Computers in ToolUniverse Connect to inspect the discovered manifest. Publishing remains an explicit step: import and live-test exactly one operation before making it public. The platform does not give marketplace callers access to tools/list or sibling tools on the same provider server.

For a server that already runs continuously, tuplatform-service install from the companion package creates a user-level launchd/systemd relay. The secret is stored in a separate 0600 file, and tuplatform-service status and uninstall provide explicit lifecycle management.

Share an Existing MCP Server

If a database or model already exposes Streamable HTTP MCP, do not wrap it again:

tu serve --forward http://127.0.0.1:9000/mcp --name "Existing model"

Use a Published Tool in ToolUniverse

Connect one public automated resource by its Discover URL or UUID:

export TU_API_KEY=...
tu connect https://connect.aiscientist.tools/discover/<tool-uuid>

Only that selected resource is saved. Calls go through the platform’s narrow /tools/call gateway; this does not grant direct relay or sibling-tool access.