Skip to main content
The MCP server ships inside proto-tools, behind an optional extra. Tools run either on this machine or on remote compute, and you choose which when you register the server. Locally, proto-tools builds each tool’s environment on first use, so no account and no setup is needed beyond the install; tools use whatever compute this machine has, so a tool that needs a GPU needs one here. Remote tools run on your own Modal account — connect it if you want remote compute, GPU models in particular.
1

Install the server

The server depends on fastmcp, which is not part of the base install:
bash
This installs a proto-tools-mcp command. Check it resolves:
bash
A plain proto-tools install does not include the server. Starting it without the extra exits with a message naming the missing dependency.
2

Connect Modal, for remote compute

Skip this step to run tools on this machine only.The server defaults to running tools on Modal, which needs a credential here:
bash
This opens a browser, and writes a token to ~/.modal.toml when you approve. Nothing is pasted, and the credential is not copied into any agent configuration.Run proto-tools doctor to check the credential, the environment, and what is deployed.Modal Setup covers accounts, environments and what it costs.
3

Register the server with your agent

The agent launches the server. You do not run it yourself.For Claude Code:
bash
For a client configured through JSON, such as Cursor or Claude Desktop:
json
Tools run on Modal by default. Add --device local to run them on this machine instead. The choice becomes the session default, and an agent can override it per call with run_on.
No credential goes into either configuration. The server reads ~/.modal.toml itself, so nothing sensitive is written into an agent’s config file.
4

Deploy the tools you want, if using Modal

Running locally, there is nothing to deploy — skip this step.A fresh Modal workspace has nothing deployed, and the agent can only run what exists. Deploy a tool before asking for it:
bash
Each deploy builds an image in your workspace and takes a few minutes.Ask the agent to call workspace_info first — it reports which workspace and environment your calls reach, and how many tools are deployed.

​
Backend requirements

The server resolves its backend at startup, and reports an unusable configuration then rather than failing on the first call. An agent can send a single call elsewhere by passing run_on to run_tool — useful for running one cheap CPU tool locally, or one GPU tool on Modal.
Tools run in a Modal workspace you own, and deploy_tool is registered. Requires Modal credentials. Deployments and runs are billed to that account, and deploy_tool asks for approval before it builds anything; declining incurs no cost.Modal Integration covers what the server does with Modal; Modal Setup covers the account, authentication, and environment.

​
Other clients

ChatGPT is not currently an option. The ChatGPT app accepts only remote MCP servers reached over HTTPS, so it cannot launch the local server described here.

​
Hosted HTTPS server

Under construction. An HTTPS version of the MCP server is in development, so an agent can use proto-tools without a local install. Tools will still run on your own Modal account.

​
Confirming the connection

The server sends the client a set of instructions for whichever backend it resolved, and the client passes them to the agent, so an agent reaches for these tools on its own without being prompted. To check a setup directly, ask for them:
  • workspace_info reports which backend the server resolved, and on Modal which workspace and environment calls go to and how many apps are deployed.
  • list_tools reports what you have deployed. Pass deployed_only=false to see the full catalog of what you could deploy.

​
Next Steps

Server Functions

Reference for each function the server exposes, with its parameters and behavior.

Cloud Inference

Deploy tools to your own Modal workspace, the backend the server uses by default.

Quickstart

The Input plus Config to Output contract the server exposes, from Python.

Tools

The catalog of bioinformatics models available through the server.