> **In short (market snapshot as of August 18, 2026).** Three different things are sold under the phrase "AI agent for 1C," and confusing them is expensive. **Developer assistants** (1C:Naparnik, Confaster, Aether Lab, Codex for 1C) write code in the Configurator; the official Naparnik is free of extra charge until October 1, 2026, and 1C has not announced the terms after that date. **Agents over your database** (42Clouds, PIX, Cinimex, Proskovya, KT.Team, OneRPA, Cloud.ru, Codex) answer an accountant's or a warehouse manager's questions about stock, settlements, and source documents; pricing runs from a one-time 12,078 RUB to a project rollout from 50,000 RUB. **Narrow AI utilities** (1C:Source Document Recognition, Entera, Epsilon, MBS) do one job — recognizing source documents, building payment orders, cleaning up item catalogs. Below: a table of sixteen solutions with prices and a snapshot date, six selection criteria, **five** ways an agent can get into 1C, and a dedicated section on **why an AI agent confidently reports the wrong number from 1C** — the core risk of this category, and one no vendor and no AI search engine writes about.

## Three markets under one query

The first thing to do before choosing is to work out which of the three markets your task belongs to. Solutions don't transfer between them: a developer assistant won't answer a question about stock levels, and a recognition utility won't explain why a supplier reconciliation doesn't balance.

**AI for the 1C developer.** Autocomplete and code generation in the built-in language, error explanations, work in 1C:EDT and the Configurator. The user is a 1C programmer.

**An AI agent over your database.** Answers questions about what's in the database: stock by warehouse, counterparty debt, turnover for a period, unposted documents, readiness to close the month. The user is an accountant, a warehouse manager, a CFO, an owner. This is the category people mean when they say "digital employee for 1C."

**A narrow AI utility around 1C.** Does one operation well: recognizes a scanned delivery note and creates the document, turns an invoice into a payment order, deduplicates the item catalog and picks classification codes. It doesn't converse and doesn't reason — it's a pipeline with AI inside.

An important shift in the summer of 2026: **the borders between these markets have blurred.** Aether Lab grew from one "Code Console" product into six and moved into both neighboring categories. MBS grew from seven catalog-data agents into five domains (master data, finance, sales, procurement, analytics) — and the natural-language report generator in their analytics domain is category two, not a utility. Codex for 1C is sold as a developer assistant, yet since version 4.7 it reads data from information bases. So "which category is this solution in" is a question about who will use it at your company, not about which section of the vendor's site it sits in.

## Comparison table: 16 solutions for 1C

| Solution | Category | What it does | How it connects | Price (August 18, 2026) |
|---|---|---|---|---|
| **Samreshuuu** | external agent over data | questions to the database, reconciliations, reports, source documents; 1C alongside your bank, EDI, CRM, and marketplaces | standard OData, database publication, a 1C user login | free to start; Pro — 2,000 RUB/month |
| **1C:Naparnik** (official) | developer assistant | autocomplete, code generation from a comment, error explanation, agent mode in 1C:EDT | built in, on an ITS subscription; requires 1C:EDT 2023.3.6+ and Java 17+ | **no extra charge until Oct 1, 2026**; later terms not announced |
| **Confaster** | developer assistant | AI inside the Configurator, built-in MCP server, BSL Language Server, analyst mode | Configurator plugin | **free** |
| **Codex for 1C** | developer assistant + data access | works with code and with information-base data; tested on Accounting, Document Flow CORP, Payroll, BIT.FINANCE, BIT:Construction | MCPBridge (Windows, .NET 4.8) plus a CFE extension; **no database publication required** | **12,078 RUB**; with Claude — 16,990 RUB |
| **Aether Lab** | developer assistant + analytics | six products: Code Console, Analyst Console, Agent Studio, a Telegram/MAX assistant, Query Console, document recognition | a data processor inside 1C | Lite free / 1,990 / 3,990 RUB per month |
| **42Clouds "Digital Employee"** | agent over data | chat with the database, diagnosing why a document won't post, questions on accounts and source documents; Accounting, Small Firm, Payroll. Since June 26, 2026, "Accountant42 Agent" alongside it creates invoices, acts, UTDs, and payment orders for the accountant to review | inside the 42Clouds cloud, 1–2 hours to connect | 7,000 / 15,000 / 30,000 / 45,000 RUB per month by request volume; on-prem by consumption |
| **PIX Operator** | agent over data | reports, documents, catalogs, registers; any configuration including custom ones; reaches email, Telegram, Bitrix24, ClickHouse, Jira | **a configuration extension**, platform 8.3.10+; MCP is its channel to external services, not to 1C | **50,000 RUB/year** BASE, one user |
| **Cinimex AI Agent for 1C** | agent over data | every database in one window, RAG over internal documents, OCR, on-prem LLM | four delivery options: standalone app, extension, the Discussions subsystem, messengers (VK Workspace, Telegram); plus 1C:Bus and MCP | price on request, enterprise |
| **AI Proskovya** | agent over data | Holding Management, Payroll, ERP, Accounting, Trade, Retail; DeepSeek, Qwen, GigaChat, on-prem; listed in the Russian software registry | extension plus an AI server plus an LLM server; SSL 3.1.8.346 or newer | **from 50,000 RUB** |
| **KT.Team AI agent** | agent over data | narrow tools plus a security gate: user, role, organization, data volume, limits, confirmation before each call, full audit log | MCP layered over HTTP services, web services, OData, or extensions | by number of processes, calculator |
| **OneRPA / OneAPA** | agent over data | low-code agents assembled from prompts, on-prem, MCP; templates for HR, accounting, sales, warehouse, support | server on your side | pricing not published |
| **Cloud.ru** | agent over data + builder | AI assistant for 1C; AI Workflows (April 9, 2026) with a 1C connector; Agents Space and GigaAgent (August 12, 2026) | Cloud.ru cloud | by calculator; 4,000 RUB starting grant |
| **1C:Source Document Recognition** (official) | utility (OCR) | the vendor's own built-in service | via ITS | **from 600 RUB/year**, 4–6 RUB per page; 250 pages a year included with PROF, CORP, and ERP |
| **Entera** | utility (OCR) | 30+ document types, counterparty checks against the state register, duplicate detection | data processor / service | "Starter" free (10 pages); Lite 76,000 RUB/year (1,200 pages per quarter); Basic 112,000 RUB/year (6,000 pages); above 10,000 pages — Custom |
| **Epsilon Workspace** | utility (treasury) | invoices into payment orders, pulling closing documents; 12 accounts across five banks: Tochka, Alfa-Bank, PSB, Uralsib, VTB | subsystems inside 1C | payback 2–3 months per subsystem, 9–12 months for the full set |
| **MBS** | utility → partly agent | five domains of seven agents each: master data, finance, sales, procurement, analytics; the analytics domain includes a natural-language report generator, variance decomposition, and group consolidation | project | on request |

Prices come from vendors' own pages and price lists as of August 18, 2026. Where a vendor publishes only "price on request" or a calculator, the table says so — we didn't invent a number to fill the cell. One warning: listicles along the lines of "Top 10 neural networks for 1C from 600 RUB" featuring products named "Neural Accountant" and "Smart 1C Consultant" are fabrications; no such products exist. Check that a product exists on the vendor's own site before treating it as an option.

## How we selected: six criteria

1. **Who the user is.** A programmer needs an assistant in the Configurator; an accountant needs an agent over the data. Getting this wrong costs more than any price difference: the free Naparnik will not answer "why doesn't the supplier reconciliation balance?"
2. **Where it installs.** A configuration extension, a data processor, a separate server, 1C:Bus, the vendor's cloud, a local bridge, or an external service over a standard protocol. This determines who updates the solution after every release of your standard configuration.
3. **Which configurations it supports.** The most common disappointment: the solution is bought, and your industry-specific or customized database isn't on the list.
4. **What it requires from your database.** Platform version, SSL version, whether the database must be published on a web server, whether you need a spare server for an LLM. That is your real rollout timeline.
5. **What happens to the data.** Does it leave for an external model, is there an on-prem option, under which account is the database read, can the agent be constrained by role permissions?
6. **Honesty of the answer.** Can the solution say "I didn't read everything" or "this total covers every organization in the database"? Nobody writes about this criterion, and it decides whether the result is usable at all. There's a dedicated section on it below.

## Market A: AI for the 1C developer

If you were looking for AI that writes code in the built-in language, the choice here is wide and mostly free.

**1C:Naparnik** is the vendor's official assistant: autocomplete, code generation from a comment, error explanation, agent mode in 1C:EDT. Through October 1, 2026 it is available at no extra charge on an ITS subscription; the product page states plainly that terms after that date will be announced separately — meaning nobody currently knows Q4 pricing. Requires 1C:EDT 2023.3.6 or later and Java 17.

**Confaster** puts AI directly in the Configurator with a built-in MCP server and BSL Language Server, works with Naparnik as well as ChatGPT, Claude, Gemini, and DeepSeek without a VPN, and includes an analyst mode for natural-language questions to the database. Free.

**Codex for 1C** arrived in July 2026 and is the most interesting thing to happen to this market in a year. It connects through an MCPBridge application on Windows plus an extension in the configuration, and runs in two modes: against a database simply open on the user's computer, and against a published one. 12,078 RUB one-time, or 16,990 RUB with Claude.

**Aether Lab** is no longer one product but six: Code Console, Analyst Console, Agent Studio, a personal assistant in Telegram and MAX, Query Console, and document recognition. Code Console tiers: Lite free, 1,990 and 3,990 RUB per month.

This segment occupies almost the entire top of search results for "neural network for 1C" — Habr articles on vibe-coding, the Infostart marketplace, the open genlab-1c/prism model benchmark. The same thing happens inside AI search engines: ask ChatGPT, Perplexity, or Alice "which AI agent should I choose for 1C" and all three answer about development by default. ChatGPT returns a ranking of Claude Code, Cursor, and Codex paired with an MCP server; Perplexity names Naparnik and local models; Alice adds PIX and MBS Group. If you're an accountant who arrived on this query, you're in the wrong place — and that isn't your fault: the market carries one name, there are three of them, and the machines that now answer questions instead of a search box mostly know about the first.

**What we do here.** We don't sell a separate "developer assistant" product, but the development side is covered by a set of skills inside the same agent: the 1C language with a BSL static analyzer, reading an XML configuration dump (object inventory, attribute types, form structure, the dependency graph, and an assessment of what a given edit will touch), the data composition system, managed forms, configuration extensions without unlocking support, external data processors, and the Configurator's batch mode. The practical point is that "why does this report calculate the wrong figure" and "what breaks if I change this attribute" go to one counterpart rather than two separate subscriptions.

The limits here are honest: the analyzer sees syntax but does not resolve the metadata of your specific configuration — a clean report means "grammatically clean," not "works"; the code isn't executed, since we don't run the platform. Binary `.cf`, `.epf`, `.dt` files, the 1C:EDT format, and ordinary (non-managed) forms aren't readable: their structure lives in a binary `Ext/Form.bin`.

## Market B: an AI agent over your data — where the real choice happens

There's plenty to compare here, and the main differences are not in model quality but in what the solution demands from your infrastructure.

**42Clouds "Digital Employee"** is the most accessible self-service option: chat with the database, diagnosis of why a document won't post, questions about accounts and source documents, 1–2 hours to connect. Tiers are 7,000 / 15,000 / 30,000 / 45,000 RUB per month by request volume, with on-prem billed by consumption. The constraint: it runs inside the 42Clouds cloud and works with standard Accounting, Small Firm Management, and Payroll. In June 2026 "Accountant42 Agent" appeared alongside it — it no longer only answers but creates invoices, acts, UTDs, payment orders, and counterparties, hands them to the accountant for review, and replies to the client in whichever channel the request arrived through. Price not published.

**PIX Operator** covers reports, documents, catalogs, and registers on any configuration including custom ones, with outbound reach into email, Telegram, Bitrix24, ClickHouse, and Jira. At 50,000 RUB a year for the BASE tier with one user, it's the clearest price on the market. One important correction to how this is often described: **inside 1C, PIX runs as a configuration extension**, platform 8.3.10 or later; MCP is its channel to external services, not its way into the database. It's also sold and implemented by Perviy Bit under their own methodology.

**Cinimex AI Agent for 1C** is an enterprise solution: every database in one window, RAG over internal documents, OCR, and the option to keep the model on-prem. There are now four delivery options — a standalone application, an extension inside 1C, integration through the Discussions subsystem, and messenger connections (VK Workspace, Telegram) — in addition to 1C:Bus and MCP. The price list is a PDF on request; this is a project sale.

**AI Proskovya** (1C PRO Consulting) is a configuration extension plus an AI server plus an LLM server, supporting Holding Management, Payroll, ERP, Accounting, Trade, and Retail, with DeepSeek, Qwen, GigaChat, and on-prem deployment. From 50,000 RUB, and listed in the Russian software registry. The requirement that's easy to miss: the standard subsystems library must be 3.1.8.346 or newer — that is, a reasonably fresh standard configuration.

**KT.Team** is an MCP server with narrow tools and, unusually, a well-documented security gate: user, role, organization, data volume, limits, and a confirmation requirement are checked before every call, and everything lands in an audit log. Their own wording captures MCP's role on this market better than anyone's: the tools "adapt to the interfaces 1C supports — HTTP services, web services, OData, or extensions." MCP doesn't replace a way into the database; it layers on top of one.

**OneRPA / OneAPA** offers low-code agents assembled from prompts, on-prem, over MCP, with ready templates for HR, accounting, sales, warehouse, and support. Pricing isn't published. **Cloud.ru** went from one assistant to three products over 2026: the assistant for 1C itself, the low-code AI Workflows with a 1C connector (April 2026), and the Agents Space with GigaAgent (August 12, 2026).

**Codex for 1C** belongs here too: since version 4.7 it reads information-base data, not just code. For this category the significance is price — 12,078 RUB one-time against 50,000 RUB a year from its nearest functional neighbor.

Three things are common to the category. First, **the entry barrier is usually a project**: a pilot, an implementation, "price on request," a separate server for the model; the exceptions are 42Clouds, PIX, and Codex. Second, **MCP has become the de facto standard** — Cinimex, KT.Team, OneRPA, Confaster, and Codex all use it, and the open catalog of MCP servers for the 1C ecosystem already lists more than fifty. Third, **1C itself still has no agent over accounting data.** The vendor has five AI services — Naparnik for 1C:EDT, Naparnik for ITS (user consultations about configurations through the support portal), Sales Forecasting, Source Document Recognition, and Speech Synthesis and Recognition. None of them answers "show me the receivables for this counterparty as of yesterday."

## Market C: narrow utilities — when an agent is overkill

**1C:Source Document Recognition** is the vendor's own service, and it's the one most often left out of the calculation: from 600 RUB a year, 4–6 RUB per page, with 250 pages a year already included in PROF, CORP, and ERP licenses, and 30 pages in Basic and Retail. Before buying separate OCR, check whether what's already included is enough.

**Entera** recognizes source documents: 30+ document types, counterparty checks against the state register, duplicate detection. The tiers have changed: a free "Starter" at 10 pages, "Lite" at 76,000 RUB a year for 1,200 pages per quarter, "Basic" at 112,000 RUB a year for 6,000 pages, and anything above 10,000 pages a year moved to an individually priced Custom plan. **Epsilon Workspace** turns invoices into payment orders and pulls in closing documents, working with twelve accounts across five banks — Tochka, Alfa-Bank, PSB, Uralsib, and VTB; stated payback is 2–3 months for a single subsystem and 9–12 months for the full set. **MBS** is no longer only master data: five domains of seven agents each, and the natural-language report generator in their analytics domain effectively moves part of the offering into category B.

The rule is simple: if the task is a single one that repeats thousands of times a month, a utility is cheaper and more accurate than an agent. An agent is for when the question is different every time.

## How an AI agent actually gets into 1C

There are five ways, and they determine almost everything else.

**A configuration extension or data processor.** Code goes inside your database. Upside: access to everything, including what the API doesn't expose. Downside: compatibility with every update of your standard configuration, and responsibility for what that code does inside your accounting perimeter. This is how Proskovya, PIX, and some Cinimex deployments work.

**1C:Bus or a corporate integration platform.** The industrial route for an enterprise with a dozen databases. Upside: a single point. Downside: it's a separate project with its own budget.

**An MCP server.** A layer that publishes a set of tools for a language model. Upside: flexible, and the set of operations can be constrained. Downside: someone has to install, update, and guard that layer — and, crucially, it still enters 1C through one of the other routes: an HTTP service, a web service, OData, or an extension.

**A local bridge to an already-open database.** New in 2026, arriving with Codex: a program on the user's computer talks to a database that is already open in 1C — no web-server publication at all. It suits file-based, local, and test databases, as well as client databases where you'll never be allowed to publish. The downside is that it works only while the computer is on and the window is open, so it can't back a scheduled routine.

**The standard OData interface.** The platform's own built-in REST protocol, available since version 8.3.5. You publish the database on a web server, enable the set of objects, create a user with the "Remote OData access" role — and an external system reads catalogs, documents, and registers with ordinary HTTP requests. Nothing is embedded in the configuration: no data processors, no extensions. Updating your standard configuration doesn't break the integration, and the load on the database is that of one more web client. In 1C:Fresh the interface is enabled for every tenant, with the object set configured by the tenant owner.

Two clarifications about OData that are easy to get wrong. First, the platform implements **version 3.0** of the specification with 1C's own extensions and limitations; no move to v4 was announced in 2026, and the two limitations below follow directly from the protocol version. Second, reports, scheduled jobs, and users are not available through this interface — if a solution promises to "read your report," it's getting in some other way.

We work through the fifth route, and it's the only one that doesn't create a new thing for you to maintain.

## Why an AI agent gets the numbers wrong in 1C

This is the core section of this article, and no vendor material contains it — and as of summer 2026 it turns out AI search engines don't contain it either: ask any of them about choosing an agent for 1C and you'll get a breakdown of models and connection methods, but nothing about why the answer might not be trustworthy. Alice goes furthest, listing "data quality in the database" among the risks — and the problem isn't data quality.

The problem with the "agent over 1C data" category isn't that it fails with an error. The problem is that **it answers successfully and incorrectly**: the platform returns 200, the number looks plausible, and you can't verify it without a second export. Here are eight mechanisms, each of which yields a silently wrong answer. We list them not for effect — this is the set of questions worth asking any vendor.

**A truncated schema arrives as a success.** Before asking anything, an agent reads the description of your database — which documents, catalogs, and registers exist in it. It's the largest response of the whole session: about two and a half megabytes for a standard configuration, tens of megabytes for CORP and industry ones. If that response is cut off midway it still arrives as a success and still looks complete, because the list of entities sits last in the file. An agent that didn't verify completeness sees an empty list and tells you: "your database has no such object." That's a claim about your database made on the basis of its own timeout.

**The schema shows composition, not access.** Role permissions aren't reflected in the database description at all. An entity in the schema means "it exists," not "I'll be shown it" — the only way to find out is to probe. An agent that plans an answer from the schema alone will confidently promise a report over data it won't be given, and won't be able to estimate how much it failed to see.

**Paged reading without a pinned order.** The platform won't return more than a thousand rows per request, so a selection is read in pages. 1C doesn't guarantee row order between requests, and if someone is posting documents in parallel, a row moves between pages: you'll read one twice and miss another entirely. The row count still looks right. The only fix is explicitly pinning the sort by reference on every page request.

**Period boundaries are inclusive.** "Turnover for Q2" with a period end of `30.06 00:00:00` loses the entire last day of the quarter, and a register snapshot "as of July 31" taken at the start of the day won't see a single movement made on the 31st and will return yesterday's value as today's. Neither case raises an error — it just produces a wrong number.

**A missing accounting dimension.** If the database holds several legal entities and the filter doesn't name the organization, "July revenue" adds them all together. 1C will happily run that query. You cannot defend such a number in front of an accountant, so the agent must either ask for the dimension or say plainly in its answer: "calculated across all organizations in the database."

**Role permissions cut the selection silently.** The platform's standard "allowed only" read mode removes rows closed off by access settings — with no error and no warning. A report over truncated data looks exactly like a complete one.

**Field names in register totals aren't what you expect.** In virtual tables the balance arrives in a field named `<Resource>Balance` and the turnover in `<Resource>Turnover`. Asking for the plain resource name returns zero — across the entire catalog, with a full warehouse, without a single error.

**Units of measure.** The quantity in a document line and in a register can be in different units: the base unit for the item, a package with a conversion factor for the line. Adding those together yields a successful answer and a wrong total.

To this add two structural limits of the third version of the protocol, worth knowing in advance. **There is no server-side aggregation in the standard interface** — groupings and totals come either from registers' virtual tables or are computed after export; trying to group on the 1C side always fails, on every configuration. And **there are no batch operations**: every write is its own transaction, and an atomic group of several documents doesn't exist.

In our agent, all of the above is handled by mechanics rather than by advice in an instruction: the database description counts as delivered only once it has been parsed in full, not when a successful response arrives; access to an entity is probed before the query plan is built, and whatever is closed off is named in one message; paged reading pins the order itself and marks the selection as truncated if it didn't finish; the completeness check sits in front of any arithmetic and halts the run rather than printing a warning; period boundaries are carried to end of day automatically; a missing accounting dimension is stated out loud; and a routing failure, insufficient permissions, and dead credentials are separated and lead to different actions instead of a universal "reconnect the integration." You can test this on your own database — ask the agent a question whose answer you already know, then ask it to name the dimension and the completeness.

## Samreshuuu: 1C alongside your bank, EDI, CRM, and marketplaces

In this table we're an external agent over your data, and there are three differences.

**Seven separate connections for seven configurations.** [Accounting](/integrations/onec-accounting), [Trade Management](/integrations/onec-trade), [Payroll](/integrations/onec-hr), [ERP](/integrations/onec-erp), [Small Firm Management](/integrations/onec-small-business), [Document Flow](/integrations/onec-document-flow) — and [1C:Enterprise (other)](/integrations/onec-custom) for an industry-specific, customized, or in-house configuration. This works because the agent doesn't know your object names in advance: its first request reads the description of your database and works out its structure — the applied names of documents and registers, added attributes, and tabular sections. For comparison: 42Clouds covers Accounting, Small Firm Management, and Payroll; Proskovya requires the standard subsystems library 3.1.8.346 or newer. Comparable breadth is now claimed by PIX (any configuration, via an extension) and Codex (any open database), so this point is no longer unique on its own — it's unique in combination with the next one.

**Nothing is installed into your database.** Only web-server publication and a user with read permissions on the objects you need. No extension, no data processor, no server for a model. Updating your standard configuration breaks nothing — and that is exactly the price everyone who enters through an extension pays.

**1C isn't the only source.** Real business questions rarely live in one system: a bank statement reconciled against 1C, an EDI reconciliation act against the database, marketplace sales against shipments. The same agent sees 1C together with your bank, EDI operator, CRM, and Wildberries and Ozon accounts — see [bank vs. 1C reconciliation](/use-cases/bank-vs-1c-reconciliation), [the EDI reconciliation act](/use-cases/edo-reconciliation-act), [overdue receivables](/use-cases/overdue-receivables), [duplicate payments](/use-cases/duplicate-payments), [the payment calendar](/use-cases/cash-gap-payment-calendar), [the closing-documents package](/use-cases/closing-documents-package), and [other 1C scenarios](/use-cases/1c). No solution in the table above does this: they live inside 1C.

Alongside the connector itself there are skills that need no database publication at all: [CommerceML parsing](/skills/commerceml_ru) for the 1C-to-website exchange, and [building UTDs and invoices in the tax authority's XML 5.03 format](/skills/fns_upd_xml_ru) for Diadoc, SBIS, Taxcom, and 1C-EDO. In simple cases a bank-to-1C reconciliation can also be done straight from a statement in the `1CClientBankExchange` format, without opening the database at all.

## Honestly about the limits

We keep this section at the same level of detail as the analysis of everyone else's limits above.

**The database must be published on a web server.** If it isn't and your admin is against it, that's the conversation to start with. The published object set is empty by default — and that is the single most common reason a first day goes badly: the credentials are valid, the database description is returned, and every object name comes back "not found." It looks like "the agent guessed the wrong name" and is fixed by a checkbox in the publication settings. We check for it on the very first request and say so plainly.

**The address matters.** A database published over `http://` inside your network is read from the computer you work on: the request is assembled on our side and leaves from your machine. While that computer is off or the application is closed, such a run waits. For round-the-clock operation from the cloud, the database needs to be published over `https://` with an external address, or we need to be deployed on your side. Exposing the 1C web client directly to the open internet is not advisable in any case — access should sit behind a VPN, a tunnel, or a reverse proxy.

**One request at a time.** A 1C license gives one connection per user: a second simultaneous request kills the first, and the whole batch of calls is lost with it. So we deliberately don't parallelize work against your database — and any vendor promising "agents for the whole department at once" runs into the same limit, just without writing about it.

**Very large configurations.** The description of a standard database weighs about two and a half megabytes and passes comfortably. For CORP configurations it can run to tens of megabytes, and that particular response has no paging — there's nothing to split. In that case the schema is captured to a file on your machine and handed to us; it's a one-time operation.

**Writes.** A write has no undo and no idempotency, so it always goes through confirmation: show what you're about to do, wait for an answer. Two limits we name directly because they're ours rather than the protocol's: we can't send the optimistic-locking header, so a write isn't protected against a concurrent edit; and on a dropped connection a write request is retried once — if the drop happened after 1C had already accepted the document, the retry creates a second one. We delete by setting the deletion mark, not by physically deleting. Start with reads, enable writes after a backup.

**When we're the wrong tool.** If the task is a single one repeating thousands of times a month — recognizing delivery notes, cleaning up the item catalog — a specialized utility from the third section will be cheaper. And if the database is file-based, sits on one computer, and there's nowhere to publish it, the local bridge from the fourth route is cheaper.

## Checklist: how to choose in 20 minutes

1. **Who will use it?** A programmer — Naparnik, Confaster, or Codex. An accountant, warehouse manager, or executive — the section on agents over data.
2. **One repeating operation, or different questions?** One — take a utility, and first check whether it's already included in your ITS subscription. Different — take an agent.
3. **Which configuration do you run?** Standard Accounting or Small Firm Management — almost anything fits. Industry-specific, customized, or in-house — ask the vendor about yours specifically, not about "1C in general."
4. **What has to be installed?** An extension in the database, a separate server, an LLM on your own hardware, a move into someone else's cloud, a bridge on a workstation — that is your timeline and your budget.
5. **Check the published object set.** If the solution goes over OData, open the publication settings and confirm the objects you need are included. An empty set is the most common cause of "nothing works" on day one.
6. **Where will the data be?** Does it leave for an external model or stay in your perimeter; is there an on-prem option; under which account is the database read, and can it be constrained by role permissions?
7. **Ask a question you already know the answer to.** On a demo, ask for something you're certain of: the balance for a specific warehouse on a specific date. Then ask which organizations it covers and whether everything was read. A solution that can't answer the second question can't be used for reporting.

## Frequently asked questions

**Does 1C have an official AI agent?**
Not one over accounting data, as of August 2026. 1C has five AI services: Naparnik for 1C:EDT (a developer assistant), Naparnik for ITS (user consultations about configurations through the support portal), Sales Forecasting, Source Document Recognition, and Speech Synthesis and Recognition. None of them answers "show me the receivables for this counterparty as of yesterday." Naparnik for development is free on an ITS subscription until October 1, 2026; terms after that date have not been announced.

**Do I have to install an extension in my database to connect AI?**
It depends on the solution. Proskovya and PIX install as extensions, Aether Lab as a data processor, Cinimex offers four options, 42Clouds requires moving into their cloud, and Codex installs a bridge on a computer plus an extension. Through the standard OData interface nothing is embedded in the configuration — you only need the database published on a web server and a user with permissions.

**Can I connect AI without publishing the database?**
Yes — as of 2026 this became possible: a local bridge works with a database simply open in your 1C. The limitation is that while the computer is off or the window is closed the agent doesn't run, so this route can't back a scheduled routine. For round-the-clock operation the database still gets published.

**Will AI work with our in-house configuration?**
Over OData, yes: the agent reads the description of your database and works out the applied names of documents, registers, and added attributes. Solutions tied to standard configurations or requiring a particular library version may not work with a customized database — that's the first question for the vendor.

**Can I avoid exposing 1C to the internet?**
Publication is required, but publishing to the open internet isn't: access can sit behind a VPN, a tunnel, or a reverse proxy, and a database inside your network is read from your own computer. Exposing the 1C web client directly is not advisable in any case — that's a security baseline, not a preference.

**Can AI damage something in the database?**
Separate reads from writes. For reading, a user with the "Read" permission on the needed objects is enough — it physically cannot damage anything. A write has no undo and no protection against a concurrent edit, and if the connection drops at the wrong moment a retry can create a second document. That's why writes go through confirmation, deletion is done with a deletion mark, and it's worth starting with reads and a backup.

**How much does an AI agent for 1C cost?**
From free (Naparnik, Confaster — but those are for developers) to an enterprise project. Reference points as of August 18, 2026: Codex for 1C — 12,078 RUB one-time, 42Clouds — 7,000–45,000 RUB per month, PIX Operator — 50,000 RUB per year, Proskovya — from 50,000 RUB, Cinimex and OneRPA — price on request. Count more than the license: a server for the model and the implementation often cost more than the subscription.

**How is an AI agent different from a data exchange or a connector?**
A connector moves data by a rigid if-then rule and doesn't answer questions. An agent receives a question in plain words, decides for itself where to look in the database, and answers with a number and its dimension. A detailed breakdown of that border is in [AI agent or Make, Zapier, n8n](/resources/ii-agent-ili-make-zapier-n8n).

**Why does AI sometimes report the wrong total from 1C?**
Because almost none of the error sources raise an error: an unfinished database description read as "the object doesn't exist," rows hidden by role permissions, an incomplete selection during paged reading, a lost last day of the period, legal entities added together, the wrong field names in register totals, mismatched units of measure. The answer arrives successful and looks right. Require the solution to name the dimension and admit incompleteness — see the section above.

---

**Last updated:** August 18, 2026.

**Sources:** 1C documentation on the standard OData interface (version 3.0, platform 8.3.5+, publication object set, register virtual tables) and [OData terms in 1C:Fresh](https://1cfresh.com/articles/data_odata); [the AI services page for the 1C platform](https://v8.1c.ru/platforma/iskusstvennyy-intellekt/) and [1C:Naparnik on the vendor portal](https://portal.1c.ru/applications/1C-Second-Pilot); [pricing for 1C:Source Document Recognition](https://v8.1c.ru/its/services/1s-raspoznavanie-pervichnykh-dokumentov/cena-1s-raspoznavanie-pervichnykh-dokumentov/); [Codex for 1C on Infostart](https://infostart.ru/marketplace/2748093/); [the open catalog of MCP servers for the 1C ecosystem](https://github.com/Untru/1c-mcp); vendors' own pages — [42Clouds Digital Employee](https://42clouds.com/ru-ru/promo/nejrosotrudnik/) and [Accountant42 Agent](https://42clouds.com/ru-ru/news/buhgalter42-agent-novyj-ii-pomoshhnik-dlya-raboty-s-1s/), [Cinimex](https://cinimex.ru/solutions/cinimeks-ii-agent-dlya-1s/), [PIX Operator](https://pix.ru/products/pix-operator/), [AI Proskovya](https://1cproconsulting.ru/proskovya), [KT.Team](https://www.kt-team.ru/product-pages/ai-agent-1c), [Confaster](https://confaster.ru/), [Aether Lab](https://aether-lab.ru/), [OneRPA](https://www.onerpa.ru/ai), [Cloud.ru](https://cloud.ru/solutions/ii-dlya-avtomatizatsii-v-1s), [Entera](https://entera.pro/cost), [MBS](https://mbsgroup.ru/ai/1c/analitika/), [Epsilon Metrics](https://blogs.epsilonmetrics.ru/ii-v-1s-dlya-buhgalterii-i-kaznachejstva/). Prices were verified on August 18, 2026; check them with the vendor on the date you connect.