autoria-mcp

autoria-mcp

An MCP server that exposes the AUTO.RIA used-cars REST API (auto.ria.com, via developers.ria.com) to AI agents — programmatic, agent-friendly access to the Ukrainian used-car market through the sanctioned API, no scraping.

The full tool surface is live — 30 tools, 7 dictionary resources, and 1 templated resource: curated search/lookup tools (single and batch listing details), paid statistics tools, thin endpoint mirrors, and browsable dictionary resources, all backed by tiered caching and a typed async client.

Features

• Curated, high-level tools that take human-friendly inputs (brand, model, region, year/price ranges) and resolve them to AUTO.RIA's numeric IDs for you.
• Paid statistics tools (AI average price, price-over-time, VIN decode) that fail fast with a clear error when AUTORIA_USER_ID is unset — no wasted quota.
• Thin endpoint mirrors for the long-tail dictionary/lookup endpoints.
• Dictionary resources — browse categories, colours, regions, etc. as addressable autoria://dict/... documents.
• Aggressive, tiered caching: large slow-changing dictionaries are cached on disk for 7 days (name→ID resolution costs no quota after the first fetch), while volatile search/statistics responses are cached briefly, in memory only.
• stdio and streamable-HTTP transports.

Install

Requires Python 3.11+. The package is distributed on PyPI; run it without a manual install using uvx:

uvx autoria-mcp

Or install into a tool environment with pipx / uv tool:

uv tool install autoria-mcp   # or: pipx install autoria-mcp

Quickstart: Claude Desktop (no coding required)

Want to use AUTO.RIA from the Claude desktop app? Follow these steps — no programming needed, just some copy-and-paste. Pick your platform: macOS · Windows.

macOS

1. Install uv (the small tool that runs autoria-mcp for you). Open the Terminal app and paste this, then press Enter:

curl -LsSf https://astral.sh/uv/install.sh | sh

(Full instructions: https://docs.astral.sh/uv/getting-started/installation/.)

2. Get a free AUTO.RIA API key at https://developers.ria.com. Keep it handy. A user_id is optional — you only need it for the paid price-statistics tools.

3. Find the full path to uvx. The desktop app can't see your Terminal's settings, so it needs the full path. In Terminal, run:

which uvx

Copy what it prints — usually /opt/homebrew/bin/uvx (Apple-Silicon Mac) or /usr/local/bin/uvx (Intel Mac).

This is the #1 reason setup fails. If you use just uvx instead of the full path, the app says it can't find it. Always paste the full path from which uvx.

4. Open Claude Desktop's config file. In the Claude app: Settings → Developer → Edit Config. This opens a file called claude_desktop_config.json (on macOS it lives at ~/Library/Application Support/Claude/claude_desktop_config.json).

5. Add the autoria server. Paste the block below. If the file already has a "mcpServers" section, add "autoria" inside it next to your other servers; otherwise paste the whole thing. Replace the two placeholder values:

{
  "mcpServers": {
    "autoria": {
      "command": "/opt/homebrew/bin/uvx",
      "args": ["autoria-mcp"],
      "env": { "AUTORIA_API_KEY": "paste-your-api-key-here" }
    }
  }
}

• command → the full path from step 3.
• AUTORIA_API_KEY → your key from step 2.
• For the paid tools (average price, VIN lookup), add "AUTORIA_USER_ID": "..." to env — put a comma after the API-key line when you do, e.g. "env": { "AUTORIA_API_KEY": "...", "AUTORIA_USER_ID": "..." }.

Save the file.

6. Fully quit and reopen Claude. Press Cmd + Q (just closing the window isn't enough), then open Claude again. The first start takes a few seconds while it downloads the tool; after that it's instant.

7. Try it. Ask Claude something like: "Use autoria to find used BMW 3 Series cars in Kyiv and show me a few with prices."

Heads-up on limits: the free AUTO.RIA key allows roughly 30 requests/hour and 1000/month — so ask focused questions rather than broad ones.

If autoria doesn't show up after restarting, the log file usually says why:

tail -50 ~/Library/Logs/Claude/mcp-server-autoria.log

Most common fixes: use the full uvx path (step 3), and make sure you fully quit Claude with Cmd + Q.

Windows

The same flow on Windows — the commands and a couple of Windows-only gotchas differ from macOS.

1. Install uv (it ships the uvx command that runs autoria-mcp). Open PowerShell and run either:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

or, if you use winget:

winget install --id=astral-sh.uv -e

Then open a new terminal window so the updated PATH takes effect. (Full instructions: https://docs.astral.sh/uv/getting-started/installation/.)

2. Get a free AUTO.RIA API key at https://developers.ria.com. Keep it handy. A user_id is optional — you only need it for the paid price-statistics tools.

3. Find the full path to uvx.exe. The desktop app can't see your terminal's PATH, so it needs the full path. In PowerShell, run:

Get-Command uvx | Select-Object -ExpandProperty Source

or in Command Prompt (cmd):

where.exe uvx

Copy what it prints — usually C:\Users\<you>\.local\bin\uvx.exe.

In PowerShell, use where.exe, not where. Plain where is an alias for Where-Object and won't find the executable.

This is the #1 reason setup fails. If you use just uvx instead of the full path, the app says it can't find it. Always paste the full path from step 3.

4. Open Claude Desktop's config file — use the in-app button. In the Claude app: Settings → Developer → Edit Config. This opens the correct claude_desktop_config.json for your install.

Don't hand-edit the %APPDATA% file. The documented location is %APPDATA%\Claude\claude_desktop_config.json (= C:\Users\<you>\AppData\Roaming\Claude\claude_desktop_config.json), but the Microsoft Store / MSIX build of Claude virtualizes that folder — it actually reads from %LOCALAPPDATA%\Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude\claude_desktop_config.json. Editing the plain %APPDATA% copy then makes your MCP servers silently fail to load. The Edit Config button always opens the right file — use it.

5. Add the autoria server. Paste the block below. If the file already has a "mcpServers" section, add "autoria" inside it next to your other servers; otherwise paste the whole thing. On Windows, backslashes in JSON must be doubled (\\) — or use forward slashes (C:/Users/you/.local/bin/uvx.exe):

{
  "mcpServers": {
    "autoria": {
      "command": "C:\\Users\\you\\.local\\bin\\uvx.exe",
      "args": ["autoria-mcp"],
      "env": { "AUTORIA_API_KEY": "paste-your-api-key-here" }
    }
  }
}

• command → the full path from step 3, with doubled backslashes.
• AUTORIA_API_KEY → your key from step 2.
• For the paid tools (average price, VIN lookup), add "AUTORIA_USER_ID": "..." to env — put a comma after the API-key line when you do, e.g. "env": { "AUTORIA_API_KEY": "...", "AUTORIA_USER_ID": "..." }.

Save the file. A ready-to-edit copy is at examples/claude_desktop_config_windows.json.

Prefer not to keep your key in the config? Omit the env block and instead put AUTORIA_API_KEY in a .env file in the working directory, or set it as a Windows user/system environment variable.

6. Fully restart Claude. Right-click the Claude icon in the system tray (near the clock) and choose Quit — just closing the window isn't enough — then open Claude again. The first start takes a few seconds while it downloads the tool; after that it's instant.

7. Try it. Ask Claude something like: "Use autoria to find used BMW 3 Series cars in Kyiv and show me a few with prices."

The same limits and most-common fixes apply: use the full uvx.exe path (step 3), edit the config via the in-app button (step 4), and fully quit from the tray (step 6). Logs are at %APPDATA%\Claude\logs\mcp-server-autoria.log.

Configuration

All settings are read from environment variables (prefix AUTORIA_) or a local .env file. Copy .env.example to .env and fill in your key.

Variable	Default	Description
AUTORIA_API_KEY	— (required for API calls)	Personal API key from https://developers.ria.com. Never logged.
AUTORIA_USER_ID	—	User id, required only by the paid POST endpoints.
AUTORIA_TRANSPORT	stdio	stdio or http (streamable-HTTP). Also --transport.
AUTORIA_HOST	127.0.0.1	Bind host for the HTTP transport.
AUTORIA_PORT	8000	Bind port for the HTTP transport.
AUTORIA_BASE_URL	https://developers.ria.com	API host (only the production host is documented).
AUTORIA_CACHE_DIR	~/.cache/autoria-mcp	On-disk dictionary cache location.
AUTORIA_CACHE_TTL	604800 (7 days)	Default dictionary cache TTL, in seconds.
AUTORIA_VOLATILE_TTL	600 (10 min)	Memory-only TTL for volatile search/statistics responses.
AUTORIA_MEMORY_CACHE_MAX	256	Max entries per in-memory cache tier (bounded LRU).
AUTORIA_MAX_RETRIES	3	Retry attempts on 429 / 5xx (backoff + full jitter).
AUTORIA_BACKOFF_BASE	0.5	Base backoff delay, in seconds.
AUTORIA_BACKOFF_CAP	8.0	Max delay for a single backoff sleep, in seconds.
AUTORIA_QUOTA_HOURLY_LIMIT	30	Assumed hourly quota; usage warns near it (never blocks).
AUTORIA_QUOTA_MONTHLY_LIMIT	1000	Assumed monthly quota; usage warns near it (never blocks).
AUTORIA_QUOTA_WARN_RATIO	0.9	Warn once usage crosses this fraction of a window limit.
AUTORIA_LOG_LEVEL	INFO	Package log level.

The API key is held in a SecretStr and never written to logs.

Running

# stdio (default) — how MCP clients launch it
autoria-mcp

# streamable-HTTP
autoria-mcp --transport http --host 127.0.0.1 --port 8000
# or
AUTORIA_TRANSPORT=http autoria-mcp

See examples/ for ready-to-paste MCP client configs (Claude Desktop and a generic stdio client).

Quota guidance

The AUTO.RIA free package is metered per API key across all RIA web services and the limits are tight:

• ~1000 requests / month
• 30 requests / hour (rolling); exceeding either returns 429 OVER_RATE_LIMIT and temporarily blocks the key.

There are no X-RateLimit-* response headers, so the client tracks usage locally (Phase 3). Treat every live call as spending scarce quota: the dictionary endpoints are cached aggressively, and search/statistics responses should not be re-fetched needlessly.

Tool catalog

Curated tools — natural inputs (names, ranges), resolved to IDs for you:

Tool	Description
search_used_cars	Search by brand/model/region/year/price/engine-volume/power/generation/etc.; returns the match count + advert ids.
get_car_details	Compact details for one advert id: price, year, mileage, structured engine volume/power, labelled body/fuel/gearbox, plus condition/risk/verification/seller/photo provenance, VIN-if-shown, masked phone, URL.
get_car_details_batch	Details for up to 50 advert ids in one call (deduped; a dead id returns a sparse entry).
lookup_brands	List passenger-car brands, or resolve one brand name to its id.
lookup_models	List a brand's models, or resolve one model name to its id.
lookup_regions	List regions (oblasts), or resolve one region name to its id.
lookup_cities	List a region's cities, or resolve one city name to its id.

Paid tools (require AUTORIA_USER_ID; fail fast with a clear error otherwise):

Tool	Description
get_average_price	AI average price + comparable listings (by params or omni_id).
get_average_price_over_periods	Monthly average-price time series.
get_params_by_vin	Decode a VIN / plate / advert id into car-parameter chips.

Thin mirrors (raw passthrough, 7-day cached): list_categories, list_all_models, list_models_grouped, list_generations, list_modifications, list_modifications_by_body, list_equipment, list_options, list_options_v2, list_colors, list_countries, list_drive_types, list_fuel_types, list_gearboxes, list_body_styles, list_body_styles_grouped, list_all_body_styles, list_bodies_by_generation, and raw_search (raw V1 search; compact {count, ids} by default, verbose=True for the full payload).

Resources: autoria://dict/{categories,colors,countries,fuel-types,gearboxes, body-styles,states} and the templated autoria://dict/models/{categoryId}/{markId}.

Health: ping — zero-quota liveness/diagnostic check.

Attribution

AUTO.RIA's API terms require a visible link back to auto.ria.com wherever you display data sourced from the API. get_car_details and the paid endpoints include a canonical per-listing deep link (url); keep it visible in whatever surface presents a listing to an end user. (search_used_cars returns only ids + a count — the Public API specifies no set-level search link, so none is invented.)

Known limitations

These are deliberate scope/behaviour choices worth knowing before you rely on them:

• Single brand/model per search. Multi-brand or multi-model search is not yet modelled by search_used_cars. For OR-style queries across several brands, drop to raw_search and pass the raw V1 params yourself.
• Search returns ids, not full listings. search_used_cars gives you advert ids and a count. Fetch per-listing detail with get_car_details (or get_car_details_batch for a whole page of ids) — each lookup spends quota, so resolve only what you need.
• Search is V1-only and silently ignores unknown params — a genuine footgun. The underlying /auto/search endpoint does not validate filter keys: a misspelt or unsupported param is dropped, and the search quietly widens to the whole site instead of erroring. The curated tool resolves names to the correct V1 ids to avoid this, but raw_search passes your params through verbatim.
• Paid POST errors arrive as HTTP 200. The paid endpoints return errors in the body (noticeType: "error") with a 200 status; the paid tools detect and surface these as tool errors.
• get_average_price_over_periods accepts only period ∈ {30, 90, 180, 365}.

Development

uv sync                     # create venv + install deps (incl. dev group)
uv run ruff check .         # lint
uv run ruff format --check .  # format check
uv run mypy                 # strict type check (src + tests)
uv run pytest               # tests (no network)
pre-commit install          # enable git hooks

The OpenAPI 3.1 description lives in openapi/autoria-used-cars.yaml.

Roadmap

• ✅ Phase 1 — OpenAPI spec, live-verified API facts.
• ✅ Phase 2 — repo scaffold, tooling, CI, packaging, runnable server.
• ✅ Phase 3 — typed async client, TTL dictionary cache, name→ID resolution, recorded-fixture tests (zero live quota).
• ✅ Phase 4 — curated + paid + thin tools, dictionary MCP resources.
• ✅ Phase 5 — docs, examples, first PyPI release (0.1.0).

Post-1.0 backlog (not scheduled): native multi-brand / multi-model search; per-endpoint cache TTLs; a cross-process quota lock (current accounting is per-process, warn-only); an npx shim for Node-based MCP clients.

License

MIT. This is an unofficial project and is not affiliated with or endorsed by RIA.com / AUTO.RIA.