science-assistant-client

High-level client for Science Assistant. It is built on mcp_proxy_adapter.client.jsonrpc_client.JsonRpcClient; it does not implement a second JSON-RPC/TLS stack.

Installation

Online:

python -m pip install science_assistant_client-<version>-py3-none-any.whl

Offline release archive:

unzip science-assistant-client-offline-<version>.zip
python -m pip install --no-index --find-links wheels wheels/science_assistant_client-<version>-py3-none-any.whl

The archive contains the client wheel, the matching mcp-proxy-adapter wheel and its dependency wheels, examples and tests.

Configuration

Environment variables:

• SCIENCE_ASSISTANT_PROXY_PROTOCOL, default https;
• SCIENCE_ASSISTANT_PROXY_HOST, default mcp-proxy.techsup.od.ua;
• SCIENCE_ASSISTANT_PROXY_PORT, default 3004;
• SCIENCE_ASSISTANT_PROXY_TOKEN_HEADER and SCIENCE_ASSISTANT_PROXY_TOKEN;
• SCIENCE_ASSISTANT_PROXY_CERT, SCIENCE_ASSISTANT_PROXY_KEY, SCIENCE_ASSISTANT_PROXY_CA;
• SCIENCE_ASSISTANT_SERVER_ID, default science-assistant-vvz;
• SCIENCE_ASSISTANT_COPY_NUMBER, default 1.

Client and server release versions must match. The first operation calls info(include_markdown=false) and raises VersionMismatchError when they differ.

Python

from science_assistant_client import ScienceAssistantClient

client = ScienceAssistantClient()
info = client.info()  # all pages are assembled automatically
client.upload_file("/mnt/data/input.fits", "incoming/input.fits")
client.download_file("exports/result.ecsv", "/mnt/data/result.ecsv")
result = client.call("astroquery_object", {"service": "simbad", "target": "M 31"})

Async usage:

from science_assistant_client import AsyncScienceAssistantClient

async with AsyncScienceAssistantClient() as client:
    result = await client.query_catalog(catalog="J/ApJ/714/25", row_limit=10)

Uploads and downloads persist sidecar state files until completion. Retry with the same paths to resume from the server-confirmed raw-byte offset. Final size and SHA-256 are mandatory.

CLI

science-assistant-client info
science-assistant-client upload /mnt/data/input.bin incoming/input.bin
science-assistant-client download incoming/input.bin /mnt/data/copy.bin
science-assistant-client call astroquery_object --params '{"service":"simbad","target":"M 31"}'

Dedicated Python 3.13 environment

The release also contains science-assistant-client-py313-linux-x86_64-<version>.zip. After extraction:

./create_venv.sh /mnt/data/science-assistant-client-venv

This creates an isolated environment and installs every dependency from the included wheelhouse without network access.

The direct JsonRpcClient transport still requires a network route to MCP Proxy. In an agent runtime where call_server is supplied as a callable bridge, pass a compatible proxy_client object to AsyncScienceAssistantClient. ChatGPT's isolated Python sandbox does not expose model tools as importable Python functions, so tool calls there remain controlled by the model runtime rather than by the venv.

Library and CLI

The distribution is a normal Python package. Import ScienceAssistantClient or AsyncScienceAssistantClient from application code, and use the installed science-assistant-client CLI for shell workflows. The package-upload CLI command sends independently checksummed parts and requests verified server-side assembly.

The standalone mcp_file_parts.py model-tool bridge is a separate release artifact. It is intentionally not imported by this package and has no network dependencies.