EU Tender MCP
# EU Tender MCP
**Live, structured European public-procurement intelligence for SMEs** — as an MCP server. Search opportunities, analyze cross-border risk, and understand EU tenders — all from inside Claude, Cursor, or any MCP client.
The value here is information an LLM **cannot know**: current tender opportunities, EU thresholds, CPV codes, and cross-border risk patterns. It's built for a consulting, services, or supply SME that wants to bid on public contracts outside its home member state.
---
## Why it's genuinely useful (and not just another wrapper)
| Thing an LLM fakes from memory | What this MCP actually does |
|---|---|
| "What's the EU threshold for services?" | Returns the real figure (€143k standard) with the category it applies to |
| "Find me tenders in Germany for software" | Returns structured, filterable tender records (country/CPV/value/deadline) |
| "Is this RFP risky to bid from abroad?" | Runs a red-flag scan (bonds, turnover floors, local-presence, short deadlines) + an honest SME fit score |
| "What CPV code is 'cleaning services'?" | Returns the CPV division with aliases |
| "What does this notice type mean?" | Explains `pin`, `cn-standard`, `results`, etc. |
Every tool is **deterministic, typed, and honest about its data source** — it never fabricates "live" opportunities when a TED API key isn't configured.
---
## Tools (7)
| Tool | What it does |
|---|---|
| `search_tenders` | Search tenders by query / country / CPV / value range / deadline |
| `market_snapshot_tool` | Activity by country & sector, typical values, urgent deadlines |
| `analyze_tender` | Red-flag scan + 0-100 SME fit score (by tender id or pasted RFP text) |
| `look_up_cpv` | Translate a description into EU CPV division codes |
| `look_up_notice_type` | Explain what a notice type (pin, cn-standard, results…) signals |
| `eu_threshold` | Whether a value is above EU mandatory-publication thresholds |
| `tender_details` | Full structured record for one tender |
---
## Install & run
```bash
python3 -m venv .venv
.venv/bin/pip install -e .
.venv/bin/pytest tests/ -q # 38 tests, incl. a real MCP stdio integration
```
### Start the server (stdio)
```bash
.venv/bin/eu-tender-mcp
# or
.venv/bin/python -m tendermcp.server
```
---
## Connect a client
**Claude Desktop** — add to your MCP config (`claude_desktop_config.json`):
`mcpServers` → see [`configs/claude-desktop.json`](configs/claude-desktop.json). Point `PYTHONPATH` at the repo absolute path.
**Claude Code / Cursor / other MCP clients** — use [`configs/mcp-config.json`](configs/mcp-config.json) and install the package (`pip install -e .`) so `python3 -m tendermcp.server` resolves.
---
## Data model
- **Default provider:** a bundled set of **clearly-labeled sample tenders** (source `bundled_sample`) so the MCP works out of the box, is demoable, and is fully testable with no key.
- **Live TED (roadmap):** register a TED API key, set `TED_API_KEY`, and implement `TEDLiveProvider._fetch()` (in `providers.py`, already stubbed). Until then, **no live data is ever fabricated** — the labeled sample is returned instead.
---
## Project layout
```
eu-tender-mcp/
├── tendermcp/
│ ├── server.py ← FastMCP server wiring all tools
│ ├── knowledge.py ← CPV codes, notice types, EU thresholds, red-flag rules
│ ├── analyzer.py ← red-flag scan + SME fit scoring (offline, deterministic)
│ ├── providers.py ← provider interface + bundled/live providers
│ ├── models.py ← typed Tender / search-params models
│ ├── intel.py ← market-snapshot aggregations
│ └── sample_data.py ← labeled demo tenders
├── configs/ ← Claude Desktop / generic MCP client configs
├── tests/ ← 38 tests including real stdio integration
├── pyproject.toml
└── README.md
```
---
## Roadmap / honest notes
- **Live TED feed** is the single biggest upgrade — it needs a registered TED/OPEN API key + account feed URL (account-specific), which only a user can supply. The provider seam is ready.
- **More red-flag rules** and **jurisdiction nuance** are planned; the analyzer is transparent on purpose (it outputs *reasons*, not false precision).
- **Not legal advice.** The analyzer is a triage aid; always verify against the full tender documents.
MIT licensed.
TDQS
Scored across 7 tools
Each tool has a distinct purpose: searching for tenders, getting market snapshots, analyzing bidder fit, looking up CPV codes, explaining notice types, checking thresholds, and fetching details. There is no overlap that would cause an agent to misselect between them. The analyze_tender tool's dual input modes are intuitively separate from the other tools.
Tool names mix patterns: most follow verb_noun (search_tenders, analyze_tender, look_up_cpv, look_up_notice_type), but market_snapshot_tool and eu_threshold use noun phrases, and tender_details is noun_noun. While readable, the inconsistent structure and the odd '_tool' suffix in market_snapshot_tool reduce predictability.
With 7 tools, the server is well-scoped and covers the core needs of EU tender discovery, analysis, and reference lookups without bloat or redundancy. Each tool clearly earns its place.
The tool surface covers search, detail retrieval, market overview, cross-border analysis, CPV translation, notice type explanation, and threshold checks—covering the main workflows. Minor gaps exist, such as no ability to fetch attached tender documents or filter tenders by date range in search, but these are edge cases that agents can often work around.