Two pure-stdlib stdio servers live here. example_server.py is only an
end-to-end fixture. protein_design/ is a deployable MCP backend adapter whose
optional scientific dependencies run only in separately configured child
environments.
The server runs as an external child process, never inside a scientific kernel. ../mcp_client.py spawns it and owns the Host-side connection; ../tools/mcp.py is what the model sees, exposing connector discovery, resource reads, and tool calls to the native control plane under the usual permission, audit, and untrusted-output policy. Both servers here are stdio only: the client's Streamable HTTP transport exists for remote connectors and is never how anything in this package is reached.
| File | Responsibility |
|---|---|
__init__.py |
Declares the bundled pure-stdlib MCP-server namespace and distinguishes the fixture from the deployable adapter. |
example_server.py |
Speaks newline-delimited MCP JSON-RPC on stdin/stdout: initialize, four sample tools (echo, now, calc, random_int), one text resource, and one parameterized summarization prompt. calc walks a restricted AST instead of calling eval. |
protein_design/ |
Nine atomic protein-design tools with explicit attempts, seeds, pinned backend/checkpoint evidence, terminal records and post-run validation; heavy model packages remain outside core. |
example_server.py as a fixture and reference. The protein-design
package implements a real MCP server and real backend command adapters, but
it is not a turnkey model distribution: the operator must separately pin and
provision every backend, checkpoint, GPU and offline boundary.@openai4s/python
instead of a machine-specific interpreter path. The MCP manager resolves the
token to the current daemon's interpreter only when it spawns the server;
it also resolves matching legacy commands at runtime, while daemon startup
rewrites those old absolute-path rows into the portable form.../mcp_client.py expects; both sides currently declare 2024-11-05.