Fifteen jurisdictions of legal research behind one MCP endpoint.

Türkiye - Netherlands - Poland - Austria - Ireland - Finland - Spain - United Kingdom - European Union - Japan - Azerbaijan - Germany - legal scholarship - signed resource contracts - GLEIF entity identity.

130 tools, one hosted endpoint, no authentication:

https://arthurlegal-mcp.fly.dev/mcp

The same server also answers at /directory/mcp with the 107 read-only tools listed in the Claude connector directory: everything except the land-registry tkgm_ and fee calculator hesap_ tools. See Claude connector directory.

Anyone with the URL can call every tool. Self-hosting? Deploy your own copy (see Indexes) and use https://<your-app>.fly.dev/mcp instead.

This is the endpoint the ArthurLegal assistant packages use. You can also connect it to any MCP client on its own (Streamable HTTP transport).

Claude (claude.ai or Claude Desktop): Settings -> Connectors -> Add custom connector. Name: ArthurLegal MCP, URL: https://arthurlegal-mcp.fly.dev/mcp. Leave the OAuth fields empty.

Claude Code:

claude mcp add --transport http arthurlegal https://arthurlegal-mcp.fly.dev/mcp

Clients that take a JSON config with remote servers (Cursor, VS Code and others):

{ "mcpServers": { "arthurlegal": { "url": "https://arthurlegal-mcp.fly.dev/mcp" } } }

Clients that only launch local (stdio) servers: bridge with npx -y mcp-remote https://arthurlegal-mcp.fly.dev/mcp as the command.

Check it is up:

curl -s -X POST https://arthurlegal-mcp.fly.dev/mcp       -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream'       -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"check","version":"1"}}}'

Then call the status tool from your client to see which jurisdictions are loaded.

The hosted endpoint is provided as is, without an uptime guarantee. Your search queries reach this server and the official upstream sources it queries, and are sent to the embeddings provider when semantic search runs. Do not put client names or other confidential facts into queries; mask documents locally first (for example with ArthurMask).

Türkiye (tr_) is the largest backend: Yargıtay/Danıştay/BAM case law, AYM, Uyuşmazlık, all legislation types, Resmî Gazete, and eight regulators (EPDK, Rekabet, SPK, BDDK, KVKK, BTK, GİB, Sigorta Tahkim) behind one tr_kurum_karari_ara interface, plus a locally indexed archive for tr_semantik_ara. Source: the arthur-tr-hukuk-mcp/ folder (formerly ArthurLegalTR/).

The archive holds the decisions' text, not just their titles: 19,498 documents, every one vectorised. EPDK's decisions before 2016 were old Word files or scans; their text was read on the crawl workstation (Word files directly, scans with OCR), and when a live document yields no text tr_kurum_karari_getir answers from the archive, with a warning where the text came from OCR. Dates are checked against each decision's own date line, so a date filter sees every Rekabet decision (10,445; before 2026-10-12 most of them carried no date and fell outside any range).

tr_resmi_gazete_tara, tr_resmi_gazete_fihrist and tr_mevzuat_ara also take a konu filter (enerji, rekabet, vergi, icra): a local, network-free topic triage over titles that finds items whose titles never mention the topic (in the labelled gazette corpus the topic word appears in only 10% of icra items). In legislation it also reads the names of the laws an omnibus act amends -- "Bazi Kanunlarda Degisiklik Yapilmasina Dair Kanun" says nothing, Law 7531 amends the Enforcement and Bankruptcy Code -- and an omnibus act whose names cannot be read is never dropped. It is a pre-filter, not a guarantee, and it was measured separately per source: gazette out-of-fold sensitivity 92-100% per topic; on 60 unseen legislation titles vergi 4/4, icra 2/2, enerji 0/3. Every response says how many items it dropped, esik=0 disables it, and status reports the model and both measurements. Do not use it for publication verification.

Official Gazette annexes (tariffs, schedules, forms) are often a separate PDF of scanned pages. tr_resmi_gazete_getir lists a page's annex links as ekler and, for a PDF, the pages without a text layer as taranmis_sayfalar. tr_resmi_gazete_ek_sayfa(url, sayfa) returns one page: its text, or the page as an MCP image block that the model reads itself. A JPEG scan goes out as the PDF's own bytes with no extra package; other scans (JBIG2, CCITT) are rendered with pypdfium2, which only the full image installs, and otherwise answer with an UYARI line. The tool is on /mcp only, not in the directory profile.

The directory listing points at a narrower profile of the same server:

https://arthurlegal-mcp.fly.dev/directory/mcp

• 107 tools, all read-only. Every tool carries a title and readOnlyHint: true (tool_titles.py), so Claude can run them without asking each time. Nothing on this endpoint writes, deletes or sends anything on your behalf.
• No land-registry tools. The tkgm_ tools stay on /mcp only; the desk shows ArthurTapu as available in the ArthurLegal install.
• The same interactive views as /mcp: the article card, the decision list and the desk (see Interactive views).
• Instructions describe the tools (in English) instead of directing Claude; the facts are the same as on /mcp.
• No account and no authentication. Add it from the directory, or as a custom connector with the URL above.

Example prompts:

1. "Quote Article 21 of the Turkish Labour Law (Law 4857) with its citation."
2. "Search Turkish Court of Cassation decisions on reinstatement (işe iade) and open the most recent one."
3. "Compare the definition of personal data in the Turkish Personal Data Protection Law (6698, Article 3) with section 3 of the UK Data Protection Act 2018."
4. "Which Act governs arbitration in Ireland? Quote the section on the arbitrator's jurisdiction."
5. "Which EU regulation lays down harmonised rules on artificial intelligence, and is it in force? Give its CELEX number."

Privacy: PRIVACY.md. Support: ertug@datsproject.io or open an issue.

On /mcp only (not on /directory/mcp), tkgm_ tools fetch single parcels live from TKGM's Parsel Sorgu -- by province, district and neighbourhood with block and parcel, by coordinates, or by place name -- and work with them: plane-projected area and edge lengths, a to-scale A4 SVG sketch, an OSM map, format conversion (DXF/KML/CSV), route-corridor intersection, 2026 title-deed fee and revolving-fund tariff lookups (every figure stored with a verbatim source quote), and a bridge from the parcel to the statutes and case law to check with the tr_ tools. Live lookups pass one rate-limited gate: at most 30 TKGM requests a minute for all users and machines together, one request in flight per machine, at least 4 s between starts, at most 3,000 a day; it backs off on 429/503 and stops on 403 (see tkgm-mcp/docs/MANIFESTO.md). The first live call answers with an approval card, so start with tkgm_baslangic. Parsel Sorgu exports vertices rounded to 5 decimals (~1.1 m), so the tools report the resulting area uncertainty instead of presenting a rounding artefact as a cadastral discrepancy. On this hosted endpoint file paths are refused (parcel text is passed in icerik) and the tools that write files, fetch elevation data or read title records are disabled; run tkgm-mcp/server.py locally over stdio for those. Source: the tkgm-mcp/ folder.

On /mcp only, three tools compute Turkish legal costs in code rather than in the model: the attorney minimum fee under the 2026 Attorney Minimum Fee Tariff (Official Gazette 03.11.2026, No. 33419), court fees under Law No. 492 tariff (1) with the 2026 amounts (Official Gazette 31.12.2025, No. 33124), and the civil procedure cost advance (Official Gazette 28.10.2025, No. 33061). Every amount sits in hesap-mcp/veri/ with a verbatim source quote; every result line carries its legal basis. Rules the official text does not settle are shown outside the total, marked for checking. No files. The only network access is a freshness check: about once a day the tools read the Official Gazette's daily index pages (at most 20 page requests a day; no request data is sent). If a new tariff publication appears, the total is withheld and the result points to it; if the index cannot be read, the total is marked as unverified. A daily GitHub job runs the same check and opens an issue for each new publication. Source: the hesap-mcp/ folder.

Every tool carries its jurisdiction as a prefix -- tr_ nl_ pl_ at_ ie_ fi_ es_ uk_ eu_ jp_ gleif_ tkgm_ hesap_ az_ de_ scholar_ contracts_. Across the underlying servers get_act means six different things and search_legislation four, so the prefix is what keeps a Spanish question from being answered with Finnish law.

Every tool also carries a readable title and a safety hint (tool_titles.py): research tools are readOnlyHint: true; the few land-registry tools that write a file when run locally are marked as not read-only and not destructive.

status reports every jurisdiction at once: which backends loaded, how many documents each has indexed, how many are vectorised, and whether semantic search is live. A backend that fails to load is announced there rather than quietly returning nothing -- "no results" and "not searched" are different answers. The German backend runs as a separate process, so status also probes it live (de_live); Azerbaijan's probe reports whether e-qanun.az answers now. On the hosted endpoint it also names the image that answered (build) and the memory in use (memory).

In Claude (web, desktop and the iOS/Android apps) results come with ArthurLegal views. The tools' JSON is unchanged, so clients without MCP Apps read exactly what they read before.

On /mcp every tool opens the same view, the desk (ui://arthurlegal/masa.html): Claude's full-screen pane shows the newest view, so a tool with a card of its own would replace the desk. Once the desk is open, each call becomes a live step on it (what was searched, how many results; the article or decision list opens under the step). In a chat without the desk the same view draws the old card for the two tools below and a one-line step for the rest. /directory/mcp keeps the two cards as before.

• tr_mevzuat_madde_getir: an article card with the law and Official Gazette details, each article's heading (with its parent headings for codes such as TBK and TMK), the text, the citation to quote with a copy button, a link to the official text and a verified / partly verified / not verified badge. An article that could not be fetched carries the UYARI line.
• tr_ictihat_ara: a decision list with chamber, docket numbers and date; each decision's text opens inside the card and can be handed to Claude for a summary.
• al_masasi: ArthurLegal Masası, a full-screen desk with legislation and citation check, case law, source status and the other ArthurLegal modules in one panel. On this endpoint the modules that run on the lawyer's computer (memory, masking, UYAP) are shown as desktop-only; on /directory/mcp land registry is shown that way too. The desk keeps its history (searches, fetched articles, answers) for 90 days in its own storage on the device, across chats and updates; nothing is written to the server.
• al_masada_basla (/mcp only): puts the question on the desk as soon as Claude starts, before the research; the steps and the answer then join it.
• al_masaya_yaz (/mcp only): while the desk is full screen the chat transcript sits behind it and the desk cannot read it, so Claude hands its answer over with this tool (question, steps, Markdown answer, sources). The host reopens the desk with the tool's input and the answer streams in as it is written; the desk then checks the answer's statutory citations against the official text. The server stores nothing and echoes nothing back. The order (start, research, hand over; "Cevap masada." in the chat) is part of this endpoint's instructions, near the top. Neither tool is on /directory/mcp, whose descriptions give no instructions; there the answer stays in the chat. The MCP screen lists connected servers and their status (Claude reports them when it opens the desk), not their tools.

The views are in arayuz/ (served as ui://arthurlegal/...). Light theme by default, dark and automatic on request; the device's system font; the only origin the sandbox allows is unpkg, for the pinned MCP Apps client (@modelcontextprotocol/ext-apps), which the browser loads only if its sha384 fingerprint matches the one in the view's import map.

Two sources stand behind the de_ prefix:

• de-eli (pinned de-eli-mcp==0.5.4, a separate FastMCP process proxied over HTTP): NeuRIS legislation and case law (/v1/rechtsprechung, seven federal courts since 2010, full text), rechtsprechung-im-internet.de, the Bundestag DIP and Open Legal Data. The aggregator reranks its search results with the same embeddings as every other jurisdiction, rebuilds decision citations from NeuRIS's German field names, returns de_recent_changes newest first and compacts DIP results.
• de-gii-mcp (de_norm_getir, de_gesetz_ara): the consolidated text of one norm from gesetze-im-internet.de. NeuRIS is still in its test phase and lacks the core codes -- on 2026-10-09 /v1/legislation?abbreviation= returned 0 for BGB, HGB, StGB, ZPO, StPO, AO, UrhG, GWB and InsO -- so de_search points there when an abbreviation is missing.

Six jurisdictions are searched from a local SQLite index (FTS5 + vectors); the rest are queried upstream and reranked in memory. The indexes are built on a workstation and baked into the image at baked/<jurisdiction>.db:

python crawl.py --embed          # in each jurisdiction directory
python build_bundle.py --target fly --out fly-app
flyctl deploy

Crawling is a maintenance task, not something a booting container does. Baking the databases in means a machine is never healthy-but-empty, and the corpus can never drift away from the code that was tested against it. start.sh still falls back to crawling when no baked index is present, and says which it used: indexes present: N/6.

A full image carries the indexes, so every full deploy uploads about 2.3 GB of build context. Most releases change only code. Dockerfile.code starts from the last full image in the app's registry and copies the code over it, so the upload is a few megabytes and a deploy takes minutes:

git worktree add --detach ../deploy origin/master   # clean checkout: no baked/
cd ../deploy
flyctl deploy . --dockerfile Dockerfile.code --depot=false --image-label <label>

Run it from a clean checkout: *.db files are ignored by git, so a fresh worktree has no baked/ and nothing large is sent. --depot=false uses Fly's own builder, which pulls the base image from the app's registry. Dockerfile.code also runs the same pinned pip install line as Dockerfile (tests/test_paketler.py keeps them equal): a new or changed Python package is fetched by the builder, so it ships with a code-only deploy too. Build a full image (Dockerfile) only when the indexes or the base Python image change, then use its tag from then on: --build-arg BASE=registry.fly.io/arthurlegal-mcp:<full-image-tag>. Files deleted from the code since the base image stay in the image until the next full build.

.github/workflows/fly-dagitim.yml runs the code-only deploy above on GitHub's machines: when the tests pass on a push to master, it deploys that commit (only if it is still the tip of master, so an older commit never lands over a newer one), then checks the live server with .github/dagitim/canli_denetim.py: every running machine is on the new image, the tool counts did not drop, the desk page matches the commit, the calculation cards answer. If a check fails it redeploys the previous image and opens an issue. If Fly's builder does not start, the build is tried once more and then made on the GitHub machine itself (.github/dagitim/dagit.sh); a failure after the build is not retried. It needs the repository secret FLY_API_TOKEN, a deploy token scoped to this app (flyctl tokens create deploy -a arthurlegal-mcp). A manual run (Actions, "Fly dağıtımı", Run workflow) redeploys master; tick "Araç sayısı bilerek azaldı" when a release removes tools on purpose. Full images are still built by hand.

hybrid, the default search mode, ranks keyword and meaning matches together, but not blindly. When the keyword channel found every word of the query, both count; when it matched only some of the words -- the usual case for a question asked in plain words -- the ranking is by meaning and the keyword matches only fill a short list. A query that is a document's identifier (an ECLI, a BOE id) lists that document first. Each response says which happened (retrieval.keyword_match, retrieval.ranking, retrieval.exact_ref). The rule comes from a 120-query known-item measurement over the six local indexes (2026-10-09). mode: "lexical" and mode: "semantic" still give either channel alone. With numpy installed (the image has it) the semantic scan is about six times faster; without it the ranking is the same.

A document holds exactly one vector (vecs.doc_id is the primary key), so changing EMBEDDINGS_MODEL does not add a second vector -- the next embedding run overwrites the old one. Until it does, queries find no vectors for the new model and semantic search reports itself off. That is deliberate: bge-m3 and voyage-4-lite are both 1024-dimensional, so mixing them would raise nothing and rank by distances computed across two unrelated vector spaces.

Without a key the server still answers, on BM25 and fuzzy matching, and every response says semantic: "off" rather than passing a keyword match off as a conceptual one.

AUDIT_LOG=1 ile açılan ağ geçidi, modelin "Madde kontrolü: … okundu" beyanını bağımsız bir kayıtla karşılaştırmak için her araç çağrısını tek satırlık bir JSON kaydı olarak HER ZAMAN standart hata (stderr) akışına yazar: fly logs stdout ile stderr'i birlikte toplar, stdio'da protokol kanalı (stdout) temiz kalır. Kayıt satırları {"olay": ile başlar; tüketiciler bu öneke göre süzer (stderr'de sunucunun başka satırları da vardır). Ayrıntı: denetim.py.

• olay: "cagri": ts, session_id, request_id, tool, params, resolved (tr_mevzuat_madde_getir: kanun/madde/ok/mulga/durum; durum arka ucun ham madde durumu, mulga yalnız durum == "repealed"; tr_mevzuat_getir: kanun/mevzuat_id), pages (sayfa, toplam, ilk-son madde), mulga, status (ok, partial, error, rate_limited, upstream_blocked, unavailable), duration_ms, sure_siniri_asildi (≥ 100 sn: Claude.ai çağrıyı keser), backend_version, returned (bütün araçlarda modele giden metnin sha256'sı ve ilk 300 karakteri), kimlikler (sonuç listelerindeki kayıtların document_id, karar_id, ref, esas_no, karar_no, citation gibi kimlik alanları, en çok 100; başlık ve metin yok).
• olay: "iptal": notifications/cancelled; çağrıyla session_id + request_id ile eşleşir.
• Maskeleme yok: params arama metni dahil argümanların olduğu gibi kopyasıdır. Gizlenen her alan ölçülemeyen bir davranış olurdu ("aradım, bulamadım" beyanı, arama sonucundan atıf). İstemcinin session_id, request_id, reason metinleri en çok 200 karakter yazılır.
• Dönen metnin tamamı değil sha256 + ilk 300 karakteri yazılır: metni kesin tanımlar, aynı parametrelerle yeniden getirilebilir, log satırını şişirmez. Kaynak sonradan değişirse eski metin geri gelmez; fark yalnız özetten görünür.
• Yalnız açıkken initialize yanıtı Mcp-Session-Id başlığı taşır (kimlik doğrulanmaz, bilinmeyen kimlik reddedilmez). Kapalıyken başlık yoktur, yanıtlar bayt bayt aynıdır.

Herkese açık uygulamada kapalı kalmalı. Yalnız ayrı test uygulamasında ve yalnız sentetik sorularla (atıf sınav kümesi): sunucu kimlik doğrulamasız, adrese gerçek dosya bilgisiyle bağlanan olursa o bilgi de kayda girer. AUDIT_LOG tanımsız ya da 1 dışında bir değerse kayıt kapalıdır.

Canlı uygulamada açılmaz: FLY_APP_NAME fly.toml'daki ad (arthurlegal-mcp) iken AUDIT_LOG=1 yok sayılır, stderr'e tek satır uyarı düşer ve sunucu kayıtsız çalışmayı sürdürür. Açıkken /directory/mcp sunulmaz, yalnız /mcp kalır: test uygulamasında dizin trafiği yoktur ve o yolun araçları sarılmadığından çağrıları kayda girmez, ölçümü yanıltırdı. PRIVACY.md'deki "writes no request logs" beyanıyla çelişmez: o beyan canlı uç noktalar (arthurlegal-mcp.fly.dev) içindir ve kayıt orada hiç açılmaz.

Test uygulaması fly.denetim.toml ile kurulur. Dosya fly.toml'ın kopyasıdır, yalnız dört satırı farklıdır: ad, boştayken durma (iki satır) ve TKGM makine sayısı. Ayrılmadığını tests/test_denetim.py sınar. Uygulama canlıyla aynı Fly organizasyonunda açılır; böylece Dockerfile.code canlının taban imajını kullanır, az-relay iç ağdan erişilir.

fly apps create arthurlegal-mcp-denetim
fly secrets set AUDIT_LOG=1 EMBEDDINGS_API_KEY=<voyage-anahtarı> -a arthurlegal-mcp-denetim
flyctl deploy . -c fly.denetim.toml --dockerfile Dockerfile.code --depot=false

Doğrulama: loglarda denetim logu AÇIK satırı görünür, /directory/mcp 404 döner. Makine boştayken durduğundan ölçümden önce fly machine start -a arthurlegal-mcp-denetim ile kaldırılır; yoksa ilk çağrı başarısız sayılır. fly logs yalnız yakın geçmişi gösterir, ölçüm süresince kayıtlar dosyaya alınır:

fly logs -a arthurlegal-mcp-denetim | grep '{"olay":' > denetim-kaydi.jsonl

Source: github.com/beerbottle90/arthurlegal-mcp · License: MIT; each backend folder carries its own copy.