# Neta MCP

**Ask an AI about any Indian MP and get answers from real parliamentary records, with sources.**

> "Who raised NEET in Parliament?" · "Give me the report card of my MP from Jorhat." ·
> "Compare Supriya Sule and Gaurav Gogoi." · "Do MPs in safe seats work less?"

Neta MCP is a [Model Context Protocol](https://modelcontextprotocol.io) server and web API
covering all **540 sitting MPs of the 18th Lok Sabha**. It joins three public sources into
one dataset:

| | |
|---|---|
| 📋 **57,717 questions** and **10,400 debate entries**, every one with its title and date | PRS Legislative Research |
| 🗓️ Attendance overall **and per session**, with national and state averages | PRS Legislative Research |
| ⚖️ Declared criminal cases, assets, liabilities and education (2024 affidavits) | MyNeta / ADR |
| 🗳️ 2024 winning margin and runner-up | Election Commission of India |

Parliament data covers 24 June 2024 to 13 August 2026. See [data/SOURCES.md](data/SOURCES.md)
for licences, the known gaps and how to rebuild the data.

## What makes it different

Most MP trackers show one number per MP. Neta lets you ask **what MPs actually talk about**
and **how the numbers relate**:

| Tool | What it answers |
|---|---|
| `search_parliament` | **Who raised a topic?** It searches all 68,000 question and debate titles. E.g. *NEET*: 56 MPs raised it, and 23 of them are from Congress. |
| `report_card` | One MP compared with the national and state averages: attendance **session by session**, which ministries they question most, their affidavit and their margin |
| `home_focus_pct` | A new measure: how often an MP's questions and debates name **their own constituency or state** (median: 15.9%) |
| `safe_seat_effect` | Do MPs who won narrowly work harder? (Finding below) |
| `compare_mps` | Two MPs side by side, showing who leads on each measure |
| `rank_mps`, `party_summary`, `state_summary`, `find_mp`, `mps_in_state` | Rankings and averages, filtered by state or party |

### First findings

- **A safe seat doesn't make MPs lazier, and a close one doesn't make them work harder.**
  The correlation between winning margin and attendance is −0.03, close to zero. MPs who won
  by less than 25,000 votes and MPs who won by more than 2.5 lakh both average 88% attendance.
- Median attendance is **90%**, but **17 MPs** attended fewer than half the sittings. Among MPs
  whose activity PRS tracks, **18** asked no questions at all.
- **218 of the 478 MPs** with affidavit data declared at least one pending criminal case.
  These are cases the MPs declared themselves, not convictions.

## Built carefully

- **Ministers aren't penalised.** Ministers (57 of them) and the Leader of Opposition don't sign
  the attendance register or ask questions. Their activity is marked *not tracked*, never 0%.
- **No wrong matches.** Affidavit and election data are attached only when the seat **and** the
  person's name both match. By-election winners are detected from their term-start date.
  For example, at Nanded the son replaced his late father, and their names share two words.
  Tests guard all of this.
- **Every answer carries its sources and caveats**, so the AI can cite them.

## How it works

```
You ask a question ──► AI assistant ──► server.py (MCP tools)     ┐
                                                                  ├──► neta/data.py (pandas) ──► data/*.csv
A website / app   ──► api.py (FastAPI, /docs)                     ┘
                                                                      ▲
                       scripts/fetch_*.py ──► scripts/build_dataset.py┘   (PRS · MyNeta · ECI)
```

| File | What it does |
|---|---|
| `neta/data.py` | All the logic, in plain pandas. Start reading here. |
| `server.py` | Exposes those functions as MCP tools for Claude and other assistants |
| `api.py` | The same functions as a web API |
| `explore.py` | Prints some findings, with no AI involved |
| `scripts/` | Downloads the sources and builds the CSVs |
| `test_data.py`, `test_api.py` | 24 automated checks |

## Setup (Windows)

```powershell
cd neta-mcp
python -m venv .venv
.\.venv\Scripts\python -m pip install -r requirements.txt
.\.venv\Scripts\python explore.py       # see the findings
.\.venv\Scripts\python -m pytest        # run the tests
```

**Use it with Claude Code:** open Claude Code inside this folder and approve the `neta` server
(configured in `.mcp.json`). Then ask, for example: *"Using the neta tools, who in Bihar raised
floods the most?"*

**Use the web API:**

```powershell
.\.venv\Scripts\python -m uvicorn api:app --reload
```

Then open <http://127.0.0.1:8000/docs> to try every endpoint in the browser.

## Roadmap

- [x] Real data from PRS, MyNeta and ECI, with sources recorded
- [x] MCP server: report cards, topic search, comparisons, the safe-seat analysis
- [x] FastAPI web API
- [ ] A simple website: "find your MP" by PIN code
- [ ] Topic trends over time (e.g. how often "unemployment" comes up each session)
- [ ] Rajya Sabha and state MLAs

## Credits and licence

Code: MIT. Parliament data: [PRS Legislative Research](https://prsindia.org/mptrack) (CC BY 4.0).
Affidavit data: [MyNeta / ADR](https://myneta.info). Election results: ECI via
[OpenCity](https://data.opencity.in). This project isn't affiliated with any of them.