SEAO Procurement Intelligence (MCP server/tool for Claude/GPT)

SEAO Procurement Intelligence (MCP server/tool for Claude/GPT)

In one sentence: you ask a question about Quebec's public security contracts in plain language ("Who held security at the Sûreté du Québec?", "What is Garda's win rate?"), an LLM like Claude or ChatGPT calls a set of analysis tools over a cleaned database of official procurement data, and you get a sourced answer with tables.

A full example of chat (in french) with Claude is available at the end of this file

1. The big picture

flowchart LR
    Q["🗣️ Question<br/><i>'Who usually wins security<br/>contracts in Montreal?'</i>"] --> L
    L["🧠 LLM<br/>Claude / ChatGPT<br/>(plans, explains)"] -->|MCP tool calls| S["🛠️ MCP server<br/>9 read-only tools"]
    S -->|SQL| D[("🗄️ seao.db<br/>SQLite")]
    D --> S
    S -->|markdown tables| L
    L --> R["📋 Answer<br/>tables, rankings, insights"]

MCP (Model Context Protocol) is the open standard that lets an AI assistant call external tools. This project is the server side: it turns a public dataset into tools an assistant can use, and the tools do the counting so the LLM does not have to guess.

Who it is for: a company (or an analyst) that wants to answer business questions about public tenders without writing SQL, e.g. pricing a bid, sizing up a competitor, or understanding a buyer.

2. The data

Of course adapting filters is necessary if you want to study another different domain instead of security services but the tool should work same

3. The pipeline, step by step

flowchart LR
    A["🌐 Open-data feed<br/>(Atom + CKAN API)"] --> B["⬇️ Fetch<br/>weekly OCDS JSON<br/>≈3.2 GB"]
    B --> C["🔍 Filter<br/>title keywords +<br/>UNSPSC codes"]
    C --> D["🧱 Normalize<br/>nested JSON →<br/>18 relational tables"]
    D --> E[("🗄️ seao.db<br/>≈7 MB")]
    E --> F["🛠️ MCP server"]

The build uses only Python's standard library. Changing the keyword list at the bottom of build_raw_json_db.py points the whole project at another sector.

Cleaning the raw data

Government open data is messy. The build step fixes this before anyone queries the data:

4. Why not just give the LLM raw SQL access?

This is the main design decision. A plain SQL connection lets the LLM write queries that run without errors and still give wrong answers. Three traps in this dataset cause that:

These rules are also written into the server's instructions, so the LLM reads them before it writes any query of its own.

5. The tools

A hybrid design: 3 flexible tools for any question, plus 6 ready-made analyses that apply the rules above.

flowchart TB
    subgraph Core["Flexible core"]
        direction LR
        A1["describe_schema<br/>data dictionary + code lookups"]
        A2["search_processes<br/>full-text + filters"]
        A3["run_sql<br/>read-only SELECT"]
    end
    subgraph Curated["Curated analyses"]
        direction LR
        B1["🎯 incumbent"] --- B2["💰 price_benchmark"]
        B3["🏢 supplier_profile"] --- B4["⚔️ head_to_head"]
        B5["🏛️ buyer_profile"] --- B6["📈 market_overview"]
    end

Example output, supplier_profile("Garda"):

# Supplier profile: GROUPE DE SÉCURITÉ GARDA SENC
matched 9 NEQs (aggregated together)
- Awards won (all methods): 341 across 320 processes, total 623,281,730 $
- Competitive win rate (open+limited): 21.9% (51 won / 233 bid on)
Most frequent rival bidders:
- NEPTUNE SECURITY SERVICES INC.: co-bid 108x, rival won 55
- TRIMAX SÉCURITÉ INC.: co-bid 54x, rival won 15

6. Safety

The server is read-only:

flowchart LR
    Q["SQL from the LLM"] --> C1{"single SELECT / WITH?"}
    C1 -->|no| X["❌ rejected"]
    C1 -->|yes| C2{"write / DDL / PRAGMA<br/>keyword?"}
    C2 -->|yes| X
    C2 -->|no| R["DB opened mode=ro<br/>auto LIMIT · 5 s timeout"] --> OK["✅ rows"]

• The database file is opened in read-only mode, so even a query that got past the checks could not modify it.
• Results are capped (automatic LIMIT), and any query running longer than 5 seconds is stopped.

7. Works with

8. Stack and numbers at a glance

Known limits

• Coverage is partial: SEAO's open data only reaches back reliably to about 2021. Older contracts are sparse, so figures are indicative, not exhaustive.
• The filter decides the scope. A security tender with an unusual title and no security product code will be missed.

Example