> ## Documentation Index
> Fetch the complete documentation index at: https://proto.evodesign.org/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Model Context Protocol
> Expose the proto-tools catalog to coding agents over MCP
```bash bash icon="terminal" theme={null}
pip install "proto-tools[mcp] @ git+https://github.com/evo-design/proto-tools.git"
```
The Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools
and data sources. `proto-tools` ships an MCP server, so an MCP-compatible agent such as Claude
Code or Cursor can discover the tool catalog, read a tool's schemas, and run it, all through
natural language.
Currently our MCP server only runs locally over stdio. Agents such as Claude Code and Cursor
are supported. A hosted version reachable over HTTPS is in active development.
## What the agent can do
An agent connected to the server works through the same contract the Python interface uses:
discover a tool, inspect its typed schemas, then call it.
`workspace_info` reports which backend calls go to, how it is configured, and whether its
credentials are present.
`list_tools` returns the catalog; `search_tools` matches a keyword against tool keys and
descriptions.
`get_tool_schema` and `get_tool_example` return a tool's input, config, and output models.
`get_tool_citation` returns the BibTeX for the method it implements.
`run_tool` executes a tool and blocks until it finishes. Large outputs are written to disk and
returned as file paths.
On Modal, `deploy_tool` provisions a tool that is not deployed yet, after asking you to
confirm the cost.
## Choosing a backend
The server runs against one backend, chosen when it starts, so listing and running agree about
what is available. The default is `modal`, and an agent can send a single call elsewhere with
`run_on`.
| Backend | Where tools run | Deployment | Requirement |
| ------- | ------------------------------ | -------------------------------------------- | ----------------------------------- |
| `modal` | A Modal workspace you own | You deploy tools yourself with `deploy_tool` | Modal credentials |
| `local` | The machine running the server | Nothing to deploy | A local GPU for tools that need one |
**`modal`** keeps model weights warm in your own workspace between calls, so repeat calls are
fast. Deploying and running bill your own Modal account, and `deploy_tool` asks you to approve
before anything is built.
**`local`** executes tools in the server's own process, so there is nothing to deploy and every
registered tool is available. Each tool builds its isolated environment and downloads its weights
on first use, so a first call can take several minutes; later calls reuse both.
## Next Steps
Install the `mcp` extra, register the server with your agent, and select a backend.
Reference for each function the server exposes, with its parameters and behavior.
The catalog of bioinformatics models the server exposes.
How the server uses Modal, and what an agent can deploy on your behalf.
