scitex-python MCP工具

Installation

```bash

Recommended — uv resolver, ~3 min (10–30× faster than pip on scitex[all])

uv pip install "scitex[all]"

Plain pip works but expect ~30–90 min — pip's resolver backtracks

heavily on the full extras set. See Installation Tips below.

pip install "scitex[all]"

> **Why uv?** `scitex[all]` pulls a large transitive set
> (numpy/pandas/torch/jax/playwright/openalex-local/sphinx-rtd-theme/…).
> pip's serial resolver walks version histories trying to satisfy
> every constraint and can spend 30+ min just downloading metadata
> before installing a single wheel. uv resolves the same set in
> parallel in 1–3 min. Install uv once with
> `pip install uv` (or `curl -LsSf https://astral.sh/uv/install.sh | sh`).

<details>
<summary><strong>Per-module extras</strong></summary>

Requires Python 3.10+. Prefix any of the above with `uv ` (e.g. `uv pip install scitex[plt,stats,scholar]`) for a 10–30× faster resolve.
</details>

<details>
<summary><strong>Installation Tips — timeouts, mirrors, <code>[all]</code> size</strong></summary>

`scitex[all]` pulls the full 33-package ecosystem plus heavy extras (playwright browsers, torch, jax, pymupdf, Apptainer/Docker integrations, etc.). With **plain pip** this takes **30–90 minutes** because pip's resolver thrashes on the transitive set; with **uv** it takes ~3 min. Recommended order of preference:

2. pip with extended timeouts (default 15s aborts mid-wheel on slow links)

pip install --timeout 600 --retries 5 "scitex[all]"

3. Install in groups if a single run keeps failing

uv pip install scitex[io,stats,plt] # core analysis layer first uv pip install scitex[scholar,writer] # research layer uv pip install scitex[audio,browser,dataset,cloud] # heavy extras last

Quick Start

<details> <summary><strong><code>@scitex.session</code> -- Reproducible Experiment Tracking</strong></summary>

One decorator gives you: auto-CLI, YAML config injection, random seed fixation, structured output, and logging.

import scitex as stx
import numpy as np

@stx.session
def main(
    data_path: str = "./data.csv",   # --data-path data.csv
    n_samples: int = 100,            # --n-samples 200
    CONFIG=stx.session.INJECTED,     # Aggregated ./config/*.yaml
    plt=stx.session.INJECTED,        # Pre-configured matplotlib
    logger=stx.session.INJECTED,     # Session logger
):
    """Analyze data. Docstring becomes --help text."""
    
    # Load
    data = stx.io.load(data_path)
    
    # Demo data
    x = np.linspace(0, 2 * np.pi, n_samples)
    y = np.sin(x) + np.random.randn(n_samples) * 0.1
    
    # FigRecipe Plot
    fig, ax = stx.plt.subplots()
    ax.plot(x, y)
    ax.set_xyt("Time", "Amplitude", "Noisy Sine Wave")
    
    # Save sine.png + sine.csv with logging message
    stx.io.save(fig, "sine.png")
    
    return 0

if __name__ == "__main__":
    main()

```bash $ python script.py --data-path experiment.csv --n-samples 200 $ python script.py --help

usage: script.py [-h] [--data-path DATA_PATH] [--n-samples N_SAMPLES]

Demo — Automated Research from Data to Manuscript

40 min, minimal human intervention — an AI agent using SciTeX completed a full research cycle: literature search, statistical analysis, publication-ready figures, a 21-page manuscript, and peer review simulation. More demos are available at https://scitex.ai/demos/.

<p align="center"> <a href="https://scitex.ai/demos/watch/scitex-automated-research/"> <img src="docs/assets/images/scitex-demo.gif" alt="SciTeX Demo" width="800"> </a> </p>

Aggregate ./config/*.yaml into a single DotDict

CONFIG = stx.io.load_configs(config_dir="./config") print(CONFIG.MODEL.hidden_size) # Dot-notation access

Every stx.io.load/save automatically records file hashes -- zero config

stx.clew.status() # {'verified': 12, 'mismatched': 0, 'missing': 0} stx.clew.chain("results/figure1.png") # Trace one file back to source data stx.clew.dag(claims=True) # Verify all manuscript claims

Configuration

cp -r .env.d.examples .env.d   # 1. Copy examples
$EDITOR .env.d/                # 2. Edit credentials
source .env.d/entry.src        # 3. Source in shell

Full configuration reference

Three Interfaces

Every capability in the SciTeX umbrella is reachable through three surfaces, so humans and AI agents share one toolkit:

The Python API is the primary surface; the CLI and MCP server expose the same logic for shells and AI agents. See the Quick Start below for runnable Python examples and the Full MCP reference.

Python-first (library API is primary)

Mixed (multiple equally-primary interfaces)

SciTeX and Research Workflow

<p align="center"> <img src="scripts/assets/workflow_out/workflow.png" alt="SciTeX Research Workflow" width="600"> </p> <p align="center"><sub><b>Figure 1.</b> SciTeX research pipeline -- from literature search to manuscript compilation, with every step cryptographically linked.</sub></p>

Architecture — Packages (3-Layer Cascade)

The 33-package ecosystem follows a strict dependency cascade: upstream imports middle imports downstream, never the reverse. Downstream apps must work standalone; the umbrella only orchestrates.

Upstream (orchestration — SOC, integration tests only)
    scitex (scitex-python), scitex-cloud
        │ imports / re-exposes
        ▼
Middle (shared infrastructure — wraps, doesn't replace)
    scitex-io, scitex-stats, scitex-app, scitex-ui, scitex-audio, scitex-dev
        │ integrates / wraps via plugin registry
        ▼
Downstream (standalone apps — own IO/GUI, unit tests)
    figrecipe, scitex-writer, scitex-scholar, scitex-clew, scitex-notebook,
    scitex-dataset, scitex-ssh, scitex-container, scitex-browser, scitex-linter,
    openalex-local, crossref-local, socialia, + utility leaves
    (scitex-{path,str,dict,logging,types,db,repro,audit,parallel,compat,gists,etc,core})

One-line contract: downstream does not know upstream exists; upstream does not duplicate downstream logic. See 01_ecosystem_01_upstream-and-downstream.md for full rules (testing, cascade, interfaces) and 01_ecosystem_02_dependency-and-version-pinning.md for dep-pinning.

Save and load -- format detected from extension.