Skip to main content

prozorro-mcp

An MCP server and a command-line tool for Prozorro, Ukraine's public procurement system. It keeps a local copy of the official open data for the regions you choose. Then you can ask an AI assistant (Claude, Cursor, VS Code Copilot and others), or the CLI on its own, things like:

  • "Which open tenders for laboratory reagents are in Lviv oblast, and when are the deadlines?"
  • "What will hospitals in my region buy next month? Show plans that are not tenders yet."
  • "Who wins most of the flour tenders here, and at what price per kilogram?"
  • "Who do we usually compete with (EDRPOU 12345678), and how often do we win against them?"
  • "Any red flags at this buyer: single bidders, one supplier taking most of the money?"

Every number is computed by code from the data, and the assistant only writes the text. No API keys are needed: the Prozorro API is public, and nothing here calls an LLM.

Ukrainian step-by-step guide for non-developers: docs/guide.uk.md.

Try it in one minute (bundled demo, no sync)

You need uv (or plain pip install prozorro-mcp):

uvx prozorro-mcp search --demo --query борошно --status any

The demo is a real snapshot of Lviv oblast for one week (5–12 October 2026): 5,479 tenders, 1,120 of them with public bids, and 1,619 upcoming plans. Contact persons were removed.

Claude Desktop: one-click extension

Download prozorro-mcp.mcpb and open it (or drag it into Settings → Extensions). Claude Desktop installs Python and the package by itself, so there is nothing else to install and no config to edit. In the install dialog, type your oblasts (for example Львівська, Волинська) and the days of history.

Your regions download in the background; until the first download finishes, Claude says the data is not ready and shows the progress. If you leave the settings empty, Claude asks which oblast and how many days to download. Every start fetches what changed since the last one. To add a region later, ask Claude: "download Volyn oblast for the last two weeks".

Connect it to other AI clients

The server runs locally over stdio. Pick your client:

Claude Code

claude mcp add prozorro -- uvx prozorro-mcp serve

Claude Desktop without the extension: Settings → Developer → Edit Config. The file is %APPDATA%\Claude\claude_desktop_config.json on Windows and ~/Library/Application Support/Claude/claude_desktop_config.json on macOS:

{
  "mcpServers": {
    "prozorro": {
      "command": "uvx",
      "args": ["prozorro-mcp", "serve"]
    }
  }
}

Cursor (~/.cursor/mcp.json) and Windsurf, Cline and other clients: the same mcpServers block as Claude Desktop.

VS Code (.vscode/mcp.json):

{
  "servers": {
    "prozorro": {
      "type": "stdio",
      "command": "uvx",
      "args": ["prozorro-mcp", "serve"]
    }
  }
}

To try it on the demo first, add "--demo" after "serve". Restart the client after editing the config.

ChatGPT and claude.ai in the browser only connect to remote servers over HTTPS. See HTTP mode.

Sync your own data

uvx prozorro-mcp sync --region Львівська --days 7
  • Regions work in any spelling: Львів, львовская, Lviv, Kyiv city, Київська. Repeat --region for several. Military units hide their address by law, so their tenders have no region. Add them with --region "без регіону".
  • First sync walks the change feed back over the period and downloads every matching tender. A busy oblast takes about 20 minutes per week of data (Lviv: about 5,500 tenders and as many plans, roughly 350 MB on disk), mostly waiting on the public API, which is used politely with 5 requests at a time and retries. --cpv 33 (or --cpv реактиви) keeps only your categories.
  • Next runs (prozorro-mcp sync with no options) fetch only what changed since the last run. Ctrl+C is safe: run the command again and it continues from where it stopped.
  • From the server: prozorro-mcp serve --sync-on-start fetches what changed in the background when the client starts it (add --region and --days to widen the coverage). Before the first sync, the data tools return no_data with a request to ask the user for an oblast and a period; during a sync, results carry sync_in_progress.
  • Status: prozorro-mcp status shows what is covered. Data lives in ~/.prozorro-mcp/data.db; set PROZORRO_MCP_DB to move it.

Use it without an AI client

The same numbers, as a table, CSV (opens in Excel with Cyrillic intact) or JSON. Install the command once with uv tool install prozorro-mcp, and then:

prozorro-mcp search --cpv реактиви --region Львівська            # open tenders, nearest deadline first
prozorro-mcp plans --query "лабораторні" --budget-min 100000    # upcoming purchases
prozorro-mcp flags --buyer 05498849                             # red flags for one buyer
prozorro-mcp watch add "реактиви Львів" --cpv 33696 --region Львівська
prozorro-mcp digest --since 24h --format csv --out digest.csv   # what is new for your saved searches

Run sync and digest every morning with Task Scheduler or cron (the guide shows how), and you have a daily list of new tenders without opening any website.

Tools

Eleven tools only read the local database. Each one declares readOnlyHint: true, destructiveHint: false, idempotentHint: true and openWorldHint: false. The twelfth, sync_data, downloads data, so it is marked readOnlyHint: false and openWorldHint: true. It exists only in the local stdio server, not in HTTP mode.

tool what it returns
coverage synced regions, period, counts, last sync; the command to add more
find_entity EDRPOU codes of buyers and suppliers by part of a name
search_tenders tenders by words, CPV, region, status, amount, deadline, publication date
get_tender one tender: items, lots, every bid, the winner, documents with direct links
market_overview count and value, competitive share, average bidders, unit-price spread, top buyers and winners
supplier_profile wins, buyers, CPVs, win rate, median discount
buyer_profile what it buys, from whom, procedure mix, single-bidder share
competitors who bids in the same tenders and the head-to-head score
upcoming_plans annual-plan items with no tender yet
red_flags the four flags below, each with evidence and a link
watch_digest new tenders for your saved searches
sync_data starts a background download of regions or a longer period; coverage shows the progress

When a question needs data that is not synced, for example another oblast or an older period, the result says so and gives the exact sync command. It does not return an empty list that looks like "nothing exists".

How the numbers are defined

  • Open tender: status active.tendering or active.enquiries, and the deadline has not passed as of the last sync. Prozorro sometimes changes the status late.
  • Competitive procedure: open, below-threshold, price quotation, competitive dialogue, framework selection and similar types. reporting (a direct contract) and negotiation are not competitive. The full list is in src/prozorro_mcp/models.py.
  • Decided tender: one that has an active award.
  • Unit price: award ÷ quantity. It is computed only for tenders with exactly one item, and grouped by the full CPV code and unit, so kilograms are never compared with tonnes. Other tenders are counted as excluded.
  • Win rate: decided tenders won ÷ decided tenders bid in. Bids become public only after the auction.
  • Discount: (expected value − award) ÷ expected value, for competitive tenders without lots.

Red flags

A flag is a reason to look closer, not proof of wrongdoing.

flag rule (defaults are parameters)
single_bidder a competitive procedure that ended with one bidder
supplier_concentration one supplier got more than 50% of a buyer's awarded money, with at least 5 awards
price_above_median unit price at least 30% above the median of at least 5 comparable tenders (same CPV and unit)
short_tender_period tendering shorter than 7 days (open), 15 (EU), 3 (below-threshold, price quotation). This is a heuristic, not a legal check: martial-law rules change often

Evals

They run without keys or network, all computed by code: prozorro-mcp eval --suite all. The latest results are in evals/RESULTS.md.

  • Flags: 35 labelled cases, including near-threshold negatives such as 29% vs 30% above the median or exactly half the money. Precision and recall are 1.0 for every flag.
  • Regions and CPV words: 94 spellings, 98.9% recognized. The one miss is shown in the results.
  • Reconcile: buyer and category totals from the analytics are checked against an independent recount over the raw JSON of the demo snapshot. 35 aggregates, 0 mismatches.

HTTP mode

prozorro-mcp serve --http --host 127.0.0.1 --port 8000      # MCP endpoint: http://127.0.0.1:8000/mcp

The same tools are served over streamable HTTP. To reach it from ChatGPT or claude.ai, put it behind a reverse proxy with HTTPS (Caddy, nginx). Version 1 has no authentication: the data is public and read-only. Add rate limits at the proxy before exposing it.

Limits

  • Local copy: you sync what you need. All of Ukraine for a month is about 100k requests to the public API, and that is not the intended local use.
  • The period means last modification: a sync covers tenders changed in the last N days. Old tenders that are still being updated come along.
  • Hidden addresses: tenders of military units have no region in Prozorro (see --region "без регіону").
  • CPV by words matches the official ДК 021 names. When the words are too broad, pass a code prefix. Tools echo the categories they used.

Development

python -m venv .venv && .venv/bin/pip install -e ".[dev]"    # Windows: .venv\Scripts\pip
pytest -q          # 197 tests, no network: recorded API responses and built tenders
ruff check .
python scripts/capture_fixtures.py    # refresh the recorded API responses
python scripts/build_demo.py          # rebuild the demo snapshot (live sync, about 20 minutes)
python scripts/build_mcpb.py          # the Claude Desktop extension (needs Node for the official packer)

The code is split by job: api (feed client) → sync → store (SQLite) → analytics (every number) → server (MCP) and cli. normalize handles regions and CPV.

Roadmap

  • v2: a price_raised_after_signing flag, for contracts whose price grew through amendments. This needs a sync of contract changes.
  • Later: a hosted HTTP instance for ChatGPT and claude.ai; Telegram or e-mail alerts as a separate service; owners of companies from the EDR register.
  • Not planned: a "chance to win" or "recommended bid". It would be a guess dressed as a fact.

Data and credits

Tender and plan data come from the public Prozorro API. ДК 021 (CPV) names and the region list come from ProzorroUKR/standards (Apache-2.0); see NOTICE.

MIT License.

Metadata

Release files for prozorro-mcp 0.1.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for prozorro-mcp 0.1.0
File Size Uploaded
prozorro_mcp-0.1.0.tar.gz 4.9 MB Details

Built distribution (wheel)

Table of built distributions (wheels) for prozorro-mcp 0.1.0
File Interpreter ABI Platform
prozorro_mcp-0.1.0-py3-none-any.whl Python 3 none any Details

Total release size: 9.8 MB

Release files / prozorro_mcp-0.1.0.tar.gz

Download URL prozorro_mcp-0.1.0.tar.gz
Size 4.9 MB
Tags Source
SHA-256 checksum
How to use checksums
02dda8d50ae7359175be42d35a4fa6dad53e9e8bcdd1d9663ecf1dfe8665f110
BLAKE2b-256 checksum
How to use checksums
0b05f0d4cd1acb9cf05c63dcb52577067ee5465474d76c31285950fbb7e2b21b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.12.10

Release files / prozorro_mcp-0.1.0-py3-none-any.whl

Download URL prozorro_mcp-0.1.0-py3-none-any.whl
Size 4.8 MB
Tags Python 3
SHA-256 checksum
How to use checksums
bb4646191752bfec4fa2540ca7804f2f43d2420e942334f973296e04e2987924
BLAKE2b-256 checksum
How to use checksums
76875c22bb82f0f701d406536056f7ec0f1df54f7b352a1f3e0279a4a8e60bc9
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.12.10

Release history Release notifications | RSS feed

0.1.1

2 release files

This release

0.1.0 This release

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page