mcp-1c-syntax-assistant

Русская версия

MCP server that gives AI agents precise answers from the 1C:Enterprise syntax reference.

An agent asks for an element by name and gets back a fixed-shape card — call string, parameters with types and requiredness, return type, execution contexts, platform version, description, example. Missing data is stated outright rather than silently omitted, because a skipped field is indistinguishable from "no such data" and the model fills the gap by guessing.

Reference content is in Russian: it is parsed from the Russian 1C help book. See Language support for what is available in English.

What an answer looks like

Real output of get_1c_element(name="Добавить", object="Массив"):

Массив.Добавить — процедура объекта Массив

Вызов: Массив.Добавить(<Значение>)
Параметры:
    Значение — Произвольный, необязательный
      Добавляемое значение. Если не указан, то будет добавлено значение типа Неопределено.

Возвращает: нет (процедура)
Доступность: тонкий клиент, веб-клиент, мобильный клиент, сервер, толстый клиент, внешнее соединение, мобильное приложение (клиент), мобильное приложение (сервер), мобильный автономный сервер
Доступно с: 8.0

Описание: Добавляет элемент в конец массива.
Примечание: При добавлении количество элементов массива увеличивается на 1.
Пример:
  Массив.Добавить("Первый");
  Массив.Добавить("Второй");

Related MCP server: 1C Buddy

Requirements

• Docker and Docker Compose v2

• 4 GB RAM free (Elasticsearch is configured for a 1 GB heap)

• The 1C syntax reference file shcntx_ru.hbk

The .hbk syntax reference file is not included. It is proprietary and ships with your licensed 1C:Enterprise installation — copy it from there. Do not redistribute it.

The repository does carry 25 pages out of the two books — 15 Russian and 10 English — under tests/fixtures/, as samples for the parser tests. Synthetic markup checks a defect once it's understood; the real pages catch what nobody anticipated — unclosed tags, pseudo-tags like <Value1>, nested font wrappers, Cyrillic in attributes. Requiring the books themselves instead — tens of megabytes, present neither in CI nor in an outside reader's checkout — would leave those cases untested. The books as such are still not redistributed here.

On Windows the file lives next to the platform binaries:

C:\Program Files\1cv8\<version>\bin\shcntx_ru.hbk

data/ is git-ignored, so the file never enters the repository by accident. If you deploy the server for other people, they get the service — not the file.

Quick start

Clone this repository, then, from its root:

# 1. Copy the reference book from your 1C installation into data/hbk/
#    Windows: C:\Program Files\1cv8\<version>\bin\shcntx_ru.hbk
cp /path/to/shcntx_ru.hbk data/hbk/

# 2. Start Elasticsearch and the MCP server
docker compose up -d

# 3. Watch the index fill up
curl http://localhost:8000/health

Indexing starts automatically on the first run and continues in the background. /health reports its progress:

{"status":"healthy","elasticsearch":true,"index_exists":true,
 "documents_count":23125,"indexing_status":"idle","indexing_active":false,
 "index_en_exists":true,"documents_count_en":23104,"version":"2.1.0"}

The server is ready when indexing_active is false and documents_count has stopped growing. index_en_exists/documents_count_en track the English book the same way — it's optional, so both fields are false/null until the English book is placed alongside the Russian one and indexed. Both containers bind to 127.0.0.1 only — see docs/DEPLOYMENT.md before exposing anything to a network.

Next: point your editor at the server — docs/CLIENT_SETUP.md.

MCP tools

Schemas, limits and the exact behaviour on ambiguous or missing names: docs/MCP_TOOLS.md.

Language support

Every tool takes a lang argument ("ru" or "en", default from the DEFAULT_HELP_LANG environment variable) that picks which book the answer comes from — the Russian index or the English one — and therefore its language. It is a separate axis from the name you pass in.

Under the default lang="ru", both languages resolve. НайтиСтроки and FindRows both find the same element, and so do Добавить and Add — the Russian reference book carries both names in every element page title (<h1>НайтиСтроки (FindRows)</h1>), and the indexer splits them into separate name_ru/name_en fields. 20 157 of the 20 159 element pages in the current index carry an English name; the two without one are structural pages, not regular elements missing a translation (a section header "Прочие процедуры и функции" and a page whose own Russian title repeats itself in place of an English one). Object names resolve too: list_1c_object_members(object= "ValueTable") and get_1c_element(name="Add", object="Array") both work, even though object pages don't print an English name in their own title the way element pages do — the server backfills it from the optional English index (see below) onto the matching Russian object page after indexing — in a single pass, right after both books are indexed. Of the 2 577 object pages, 2 555 (99%) have picked up an English name this way, and all 389 constructor pages did too; the remaining 22 objects answer only to their Russian name, and there is nothing to fix on this side: 21 of those pages do not exist in the English book at all (it holds 23 104 pages against 23 125 in the Russian one), and the 22nd is titled there with the book's own internal page id rather than a name, which the backfill refuses to take. Either way, the card itself — description, parameters, availability, example — is Russian, because that is the only language the Russian book carries them in.

lang="en" is a different thing: a genuinely English answer, end to end, from the optional second book (shcntx_root.hbk — see docs/CONFIGURATION.md), not a translation of the Russian one. It requires that book to be indexed, and it requires an English name — passing a Russian one is refused outright rather than silently searched for in the Russian index under the guise of an English answer (see below). Real output of get_1c_element(name="Add", object="Array", lang="en") against the current English index (23 104 documents):

Array.Add — procedure of Array

Call: Array.Add(<Value>)
Parameters:
    Value — Arbitrary, optional
      Added value. If not specified, a value of Undefined type will be added.

Returns: nothing (procedure)
Availability: thin client, web-client, mobile client, server, thick client, external connection, mobile application (client), mobile application (server), mobile standalone server
Available since: 8.0

Description: Adds an element to the end of the array.
Note: When an element is added, the number of elements in the array is increased by 1.
Example:
  Array.Add("First");
  Array.Add("Second");

A call that mixes languages the wrong way — a Cyrillic name with lang="en", or any lang="en" call before the English book has been indexed — gets an explained refusal instead of a silent empty answer. The reverse never happens: lang="ru" accepts an English name freely, because the Russian book carries both. Exact wording of all three refusal cases: docs/MCP_TOOLS.md.

Differences from the upstream project

Based on Antonio1C/1c-syntax-helper-mcp (MIT, as declared in its README), which contributed the FastAPI service layout, .hbk extraction through 7-Zip and Elasticsearch indexing. What changed:

• The element card is a contract, not free text. A fixed set of fields is always printed, and absent data is labelled (Доступность: в справке не указана, Примеров в справке нет) instead of being dropped.

• Three MCP tools instead of one, each with an explicit JSON schema, so an agent picks a tool by purpose rather than by guessing arguments.

• Disambiguation instead of a silent pick. Добавить occurs on 197 pages; the server returns an ordered candidate list and asks for the object, rather than returning an arbitrary one of them as the answer.

• Call variants, real parameter requiredness and property value types are parsed out of the help HTML.

• A reproducible search-quality measurement — scripts/eval_search.py builds its ground truth from the index itself and reports hit rates.

Full attribution and the complete list of changes: NOTICE.

Roadmap

• Reference for the 1C language, the query language and the DCS expression language (shlang, shquery, shclang). These books use a different page format and need a separate parser.

License

MIT — see LICENSE. Attribution of the upstream project and the list of changes made here: NOTICE. The upstream project is Antonio1C/1c-syntax-helper-mcp.

The licence covers this source code only. It does not cover the proprietary 1C:Enterprise syntax reference — neither the .hbk files, which are not part of this repository, nor the 25 sample pages under tests/fixtures/ that the parser tests are built on (see NOTICE).

Documentation

All four documents are in Russian.

• docs/CLIENT_SETUP.md — connecting VS Code, Claude Code and other MCP clients.

• docs/MCP_TOOLS.md — tool schemas, limits, card format, behaviour on ambiguous and missing names.

• docs/CONFIGURATION.md — environment variables, reindexing, replacing the reference file.

• docs/DEPLOYMENT.md — Windows Server and Linux ARM64, multi-architecture images, HTTP endpoints, exposing the service safely.

Contributing to the test suite: tests/README.md.