Proto Client

proto-client is the official Python SDK for the Proto project. Run any of 80+ bioinformatics tools from proto-tools, submit optimization runs built with proto-language, stream logs, and download results, all from a few lines of typed Python.

The SDK provides a synchronous ProtoClient and an asynchronous AsyncProtoClient with the same surface, fully type-checked responses, and transport-level retries. It also includes an MCP server, so Claude, Cursor, VS Code Copilot, and any other MCP-compatible agent can call the same APIs through natural language.

Installation

All you need is Python 3.10+ and pip:

pip install proto-client

The MCP server is an optional extra:

pip install proto-client[mcp]

Quickstart

Set PROTO_API_KEY in your environment (or pass api_key= explicitly), then:

from proto_client import ProtoClient

client = ProtoClient()

# Run a bioinformatics tool and poll to completion (tools API)
job = client.tools.run("esmfold-prediction", {"sequences": ["MKTL"]})

# Submit an optimization run and poll to completion (runs API)
run = client.runs.run(program_data={...})

Async

Every namespace has an identical async surface. await the calls and use the client as an async context manager:

from proto_client import AsyncProtoClient

async with AsyncProtoClient() as client:
    run = await client.runs.create(program_data={...})
    status = await client.runs.get(run.run_id)

Working with output assets

Large cloud outputs (structures, logits, PAE matrices, embeddings) come back as AssetRef objects rather than inline JSON; the client.assets namespace downloads, decodes, or streams them to disk on demand. See Working with output assets for the methods, MIME decoding, and examples.

Exporting a run's results

There are two export routes, depending on where the program ran:

The two cover non-overlapping data: the local route needs the live Program object (because seq.structure / seq.logits only exist client-side), the server route needs a run_id (because the results live in the API database). Pick by which one you have.

# Local Program (proto-language installed alongside proto-client)
program = Program(...)
program.run()
client.export_program(program, "out/")

# Server-side Run
run = client.runs.run(program_data=...)  # creates and polls to terminal status
client.runs.export(run.id, "out.zip")

For a single asset rather than the whole bundle, use client.assets.download(ref, path). Both export routes have async equivalents: await client.export_program(...) and await client.runs.export(...).

Command line

A minimal proto-client CLI submits tools and optimization runs from a shell or CI, then writes the result JSON and downloads any output assets:

proto-client me                                                    # workspace and credits
proto-client tools run esmfold-prediction --inputs in.json --assets ./out
proto-client runs submit program.json --wait --export out.zip

See the CLI guide for the full command reference.

Using with AI agents (MCP)

proto-client includes an MCP server so Claude, Claude Desktop, Cursor, VS Code Copilot, Codex, Gemini, and any other MCP-compatible agent can call the Proto Bio APIs through natural language. Connect to the hosted endpoint (nothing to install).

See the MCP user guide for per-agent connection snippets, the tool/prompt/resource surface, and a guided walkthrough.