MCP-Powered Scientific Assistant 🚀

Bring Claude 3 or GPT-4o together with any Model Context Protocol (MCP) servers you like, to build a local assistant that can:

• search, read and summarise research papers
• call custom tools (e.g., a Lucene index, Google Drive, Fetch…)
• cross-reference findings and help with systematic reviews

---

✨ Features

Capability	Details
Provider switch	provider=openai or provider=claude per request
Tool calling	Works with the official mcp SDK – any server that follows the spec
Streaming replies	FastAPI + Uvicorn StreamingResponse
Local-first	No external costs beyond your chosen LLM model; MCP servers can run on your own machine

---

🛠️ Prerequisites

• Python ≥ 3.10
• Valid API keys for OpenAI or Anthropic Claude
  (put them in .env; an example file is provided)
• One or more MCP servers running.

---

🚀 Quick start

# 1. Clone
git clone https://github.com/dantebarross/science_mcp.git
cd science_mcp

# 2. Python environment
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# 3. Configure secrets & MCP servers
cp .env.example .env
#   ↳ paste your own keys and server URLs

# 4. Run
dos2unix .env run_local.sh # If you are on WSL
chmod +x run_local.sh
./run_local.sh

Now open another terminal:

```bash
curl -N http://localhost:8000/chat \
     -H "Content-Type: application/json" \
     -d '{"message": "Search for a 2024 tDCS-EEG meta-analysis and summarise it.",
          "provider": "openai"}'

---

🧩 Adding MCP servers

This assistant can connect to any Model Context Protocol (MCP) server.

We recommend starting with three powerful local tools:

MCP Server CLI	Description	Install & Run	Default Port	Example Tools
mcp-server-fetch	Fetches any URL and returns cleaned Markdown (great for PubMed/Arxiv abstracts).	pip install mcp-server-fetch FASTMCP_PORT=9100 mcp-server-fetch --transport sse	9100	fetch(url)
minima	Indexes & vector-searches a local folder of PDFs/MD/TXT with RAG support (semantic search).	cd mcp-servers/minima/mcp-server && pip install -e . minima --dir ~/papers --transport sse --port 9200	9200	search_docs(q), read_doc(path)
mcp-server-filesystem	Sandboxed file-system access to read/search any file under a root directory.	pip install mcp-server-filesystem-python FASTMCP_PORT=9300 python -m filesystem.server ~/papers --transport sse	9300	read_file(path), search_files(q)

For minima's mcp server installation, you should install the requirements

cd mcp-servers/minima/mcp-server
pip install -e .

Run all three with:

dos2unix run_servers.sh # If you are on WSL
chmod +x run_servers.sh
./run_servers.sh # MCP tools on ports 9100 (fetch), 9200 (minima/search_docs), 9300 (filesystem)

---

## 🩺 Troubleshooting

* "`No MCP client for …`" → URL typo or server not running.
* `401` from Anthropic/OpenAI → check your keys / billing.
* Slow answers -> try Claude "Haiku" or GPT-3.5 for speed.

---

## 📜 License

This template is MIT-licensed. Feel free to adapt for research or production.