Integrated and sponsored by InfMind, an overseas AI-for-science compute platform. SciAgent-Thermal runs as one of InfMind's reference agents for multi-tool scientific-computing workflows — huge thanks to the InfMind team for the compute credits and platform integration.

SciAgent is an AI Agent framework that solves microchannel heat sink design problems from a single natural-language prompt. It follows a Prompt–Tool–Workflow three-layer architecture, exposes its scientific tools over the Model Context Protocol (MCP), orchestrates multi-tool collaboration via a Workflow layer, and executes model-generated code inside an AST whitelist + subprocess Sandbox.

• Benchmark: 10-task microchannel heat-sink evaluation set
• Pass rate: 10/10 (100%)
• Average score: 88.8 / 100

┌────────────────────────────────────────────────────────────────┐
│                         SciAgent                               │
│                                                                │
│   ┌─────────────┐   ┌──────────────┐   ┌──────────────────┐    │
│   │   Prompt    │   │   Workflow   │   │      Tools       │    │
│   │   layer     │──▶│  ReAct loop  │◀─▶│ (via MCP server) │    │
│   │ system.py   │   │   react.py   │   │                  │    │
│   └─────────────┘   └──────┬───────┘   └──────┬───────────┘    │
│                            │                  │                │
│                            ▼                  ▼                │
│                      ┌──────────┐      ┌──────────────┐        │
│                      │   LLM    │      │   Sandbox    │        │
│                      │ llm.py   │      │ ast_whitelist│        │
│                      └──────────┘      │ + subprocess │        │
│                                        └──────────────┘        │
└────────────────────────────────────────────────────────────────┘

sciagent/mcp_server/ publishes the entire Tool layer over the Model Context Protocol (JSON-RPC 2.0 over stdio, initialize / tools/list / tools/call). The same tool set is usable from any MCP host (Claude Desktop, Claude Code, Cursor, …):

python -m sciagent.mcp_server

sciagent/sandbox/ast_whitelist.py performs static analysis on every code snippet before execution:

• Imports must be in a whitelist (numpy, scipy, matplotlib, deap, pandas, stdlib math/stats).
• Forbidden names: eval, exec, compile, __import__, open, input.
• Forbidden attr calls: os.system, subprocess.*, socket.*, network libs.
• Dangerous dunder access (__class__, __globals__, …) is blocked.

Code that passes the static check is then run in an isolated subprocess with a 60-second timeout, so a crash or infinite loop in generated code cannot take down the host process.

Benchmark: 10 microchannel-heat-sink tasks, covering multi-step reasoning, parameter calculations, and tool-call chains.

Dimensions covered:

• Heat-transfer fundamentals (fin efficiency, Nu=3.66 verification)
• Fluid mechanics (hydraulic diameter, friction factor)
• Microchannel thermal analysis (pressure drop, thermal-resistance network)
• Parametric study (Nu vs aspect ratio)
• Single- and multi-objective optimization (NSGA-II)
• Comprehensive design

See eval_reports/BENCHMARK_README.md and the latest report eval_reports/eval_report_20260409_212803.md.

pip install -r requirements.txt
cp .env.example .env   # put your DEEPSEEK_API_KEY inside

# CLI chat
python main.py

# Run the full benchmark
python evaluate.py

# Boot the MCP server (for Claude Desktop / Cursor / ...)
python -m sciagent.mcp_server

sciagent_2/
├── sciagent/
│   ├── prompts/          # Prompt layer
│   ├── tools/            # Tool layer (7 tools)
│   ├── workflow/         # Workflow layer (ReAct)
│   ├── sandbox/          # AST whitelist + subprocess isolation
│   ├── mcp_server/       # MCP protocol server
│   ├── llm.py            # Unified DeepSeek / Claude / Mock wrapper
│   └── agent.py          # Thin facade composing the three layers
├── benchmark.json        # 10-task evaluation set
├── evaluate.py           # Benchmark runner & scorer
├── main.py               # CLI entry point
└── eval_reports/         # Auto-generated evaluation reports

• InfMind — overseas AI4S compute platform. SciAgent-Thermal is integrated into InfMind as a reference scientific-computing agent; the project's benchmark runs and MCP-server demos are powered by InfMind's compute credits. Thanks to the InfMind team for the sponsorship and for pushing AI-for-science tooling forward.
• DEAP — NSGA-II reference implementation.
• Anthropic's Model Context Protocol — the standard that lets this project's tools plug into any MCP-capable host.

MIT