If your team customizes 1C in 1C:EDT and you need a fast start without your own infrastructure, start with 1C:Naparnik: it is built into the development environment, runs on 1C’s servers, and was free at the time of review. A custom AI assistant makes sense when you need what Naparnik does not promise: operation inside your own environment with a chosen model, search across internal policies and past customizations, and answers for end users and consultants, not just developers. An October comparison on Habr showed that a self-built assistant on an open stack answered some real questions no worse than the vendor solution, but noticeably more slowly. The best choice is made not from reviews, but from a pilot on 20–30 of your own tasks.
Prepared by AIrassvet from public sources, verified on October 12, 2026. It is intended for IT and 1C department leaders, business owners with in-house 1C developers, and 1C partners who are deciding whether the vendor’s AI assistant is enough or whether they need their own. We did not test 1C:Naparnik and did not conduct our own comparison: the service description was taken from official 1C pages, and the comparison results came from the author’s post on Habr. AIrassvet is not affiliated with 1C or the authors of the referenced materials; the section at the end of the article describes the publisher’s services.
Contents
- Quick Take: How to Choose
- What 1C:Naparnik Is
- How Much 1C:Naparnik Costs and Until When It Is Free
- How a Custom AI Assistant for 1C Is Set Up
- What the Habr Comparison Showed
- Comparison by Criteria
- When to Choose Naparnik, Your Own Assistant, or Both
- Data and Security: What Leaves the Perimeter
- How to Compare the Options on Your Own Tasks
- Common Mistakes
- Frequently Asked Questions
- Primary Sources and Verification Method
Quick Take: How to Choose
- Naparnik is a sensible starting point for a team that already works in 1C:EDT: setup takes installing a plugin and entering a key, with no separate servers required.
- A custom assistant is needed when perimeter, model, and knowledge sources matter: code and questions must not leave even the vendor’s servers, a specific model is required, or answers must be based on your policies, tasks, and customization history.
- For 1C users and consultants Naparnik is not presented by the vendor as a target audience: on the official pages we reviewed, only a developer-focused version is described.
- Quality depends on context more than on the model name: in the Habr comparison, the open stack won thanks to hybrid search across standard configurations and web search, while losing on speed.
- Compare on your own tasks: 20–30 real questions, two independent reviews, and pre-defined criteria that specify whether the answer can be used as is.
- A free period is not a strategy: the official pages list two different end dates, and the post-free price has not been announced.
What 1C:Naparnik Is
1C:Naparnik is an AI assistant from 1C for developers, built into the 1C:EDT development environment. On the official service page it is described as an assistant that works in agent mode: it takes a task in natural language, searches information in the project, reads files, plans subtasks, and makes code changes.
According to 1C, Naparnik can:
- search the project: by metadata objects, by meaning (vector search), and by files with masks;
- change code: targeted edits, adding procedures and functions, refactoring with 1C syntax validation;
- fix errors: read 1C:EDT error and warning markers and run bsl-check;
- edit metadata: create and modify catalogs, documents, registers, enumerations, reports, and data processors, and add attributes;
- work with Git: commit history, version comparison, analysis of local changes;
- search ITS and BSP documentation, including version comparisons;
- run ready-made scenarios in chat: explain code, perform reviews, write comments for procedures, prepare a commit message.
Separate documentation describes four security modes: "Confirm changes" (default: reads are automatic, edits require confirmation), "Confirm everything," "No confirmations," and "Read only." Risky Git operations—push, reset --hard, and branch deletion—according to the vendor’s description, always require explicit approval.
On the ITS portal the service is called "1C:Naparnik for Developers." It lists the requirements: an account on the 1C:ITS portal, an activation key, 1C:EDT version 2023.3.6 or later, Java 17 or later , and a stable internet connection. One key is issued to one developer, under one ITS account — up to 100 keys; if more are needed, you are advised to contact support. The code.1c.ai page states that Naparnik is included with 1C:EDT 2025 and later, while earlier versions use a plugin installation.
What we did not find in the official description: work in Configurator, a version for accountants and other application users, deployment on the customer’s servers, or the name of the model powering the service. That does not mean such plans do not exist—it only means that as of the verification date, the vendor had not described them.
How Much 1C:Naparnik Costs and Until When It Is Free
As of the verification date, the service is free, but the official sources cite different end dates for the free period:
| Source | What it says | Verified |
|---|---|---|
| code.1c.ai | "Free until February 1, 2027" | 10/12/2026 |
| ITS portal | no additional charge until December 1, 2026; any change in terms will be announced separately | 10/12/2026 |
The price after the free period is not specified in either source. For your 2027 budget, that means one thing: plan for a scenario in which Naparnik becomes paid, and confirm the terms with your 1C partner or in your ITS account. For more on planning AI expenses, see the article on your AI budget for 2027.
How a Custom AI Assistant for 1C Works
A custom AI assistant for 1C is a language model connected to search across your configurations and documents, plus a set of tools, all orchestrated by an agentic wrapper. The model can be cloud-based or run on your own servers; the value doesn’t come from the model alone, but from the context the assistant finds before answering.
A good reference point is the stack described by the author in a Habr post from October 10:
| Component | What the author used | Why it’s needed |
|---|---|---|
| Agentic wrapper | Deep Agents on LangGraph | task planning, delegation to subagents with isolated context, history compression |
| Models | models at the level of DeepSeek V4 Flash, GPT-6.1 Sol, GLM-5.3 Flash | response and code generation |
| Full-text search | PostgreSQL FTS with a Russian dictionary | exact matches: object names, procedures, attributes |
| Vector search | pgvector with an HNSW index | semantic search when the question is phrased in your own words |
| Result merging | Reciprocal Rank Fusion | one list combining results from two search methods |
| Codebase | exports from 5 typical configurations (about 236,000 files and 193,000 directories) and a custom metadata parser | answers at the level of specific modules and objects |
| Web search | Yandex API | practical solutions not found in the official documentation |
Hybrid search — combining full-text and vector search — is not a luxury here. 1C contains many exact identifiers: module, register, and procedure names. Vector search looks for semantically similar content and can miss the exact name; full-text search is the opposite. For more on why assistants make mistakes with imprecise context, see the piece on RAG accuracy and hallucinations.
For a company, three additional sources are usually added to this setup — sources that are missing from both standard exports and the vendor’s documentation:
- your customizations: export of your own configuration and extensions;
- internal knowledge: policies, user instructions, process documentation, answers to frequently asked questions;
- task history: closed tickets and descriptions of past customizations from the task tracker.
If the assistant needs not only to answer but also to work with database data — for example, find documents or check inventory balances — you need a separate secure channel for access to 1C. We covered how to set that up in an article on how to connect AI agents to 1C.
What the Habr comparison showed
The author compared their assistant (“Agent”) with 1C:Partner on five questions from real-world 1C developer work. The evaluation was qualitative: the author checked whether the answer named specific modules, provided code, and outlined verification steps. There were no numeric scores in the article.
| Question | Custom assistant answer (per the author’s description) | Partner answer (per the author’s description) |
|---|---|---|
| Setting up the Technology Journal for managed lock timeouts | journal configuration with TLOCK, TTIMEOUT, TDEADLOCK events and a 1,000,000 µs threshold, with an explanation of how to find the blocking connection | configuration with EXCP, TLOCK, TTIMEOUT events, with emphasis on collecting data for investigation |
| Connecting a custom directory to BSP object versioning | BSP steps with code and the "WriteObjectVersion" subscription | description of a different mechanism — the platform data history — without integration with BSP |
| Preventing a posting reversal if batches have already been written off | event handler code with a query and a note about managed locks | answer stating that the sources do not contain a specific solution |
| Enabling versioning for attached files | explanation of the "Keep Versions" flag and code | advice to add attributes based on BSP 2.4 recommendations without explaining how to enable it in user mode |
| Why there is inventory balance in UT 11.5 without a record in the cost register | six possible reasons and a diagnostic sequence | two reasons and month-end closing |
The author’s conclusion: a top-tier model with good orchestration delivers results “at least no worse” than a specialized solution, and the barrier to entry has dropped significantly. Important caveats from the author: their assistant runs significantly slower because it spends time searching, reading sources, and making additional model calls; development costs were not included in the comparison; the sample was five questions, and the evaluation was expert-based.
How to interpret this result. Five questions do not prove that one tool is better than another overall. But the comparison clearly shows the mechanism: on questions whose answers are not in the official documentation, the winner is the one with broader and more precise context. Partner relies on the project, ITS documentation, and BSP; the author’s self-built assistant also uses full typical configurations and web search.
Comparison by criteria
| Criterion | 1C:Partner | Custom AI assistant |
|---|---|---|
| For whom | 1C:EDT developers | developers, consultants, users — depending on how you configure it |
| Where it runs | 1C:EDT plugin | IDE, web chat, messenger, support portal — your choice |
| Where the computations run | 1C company servers in Russia, according to the vendor | cloud or your own servers; in an on-premise model, data never leaves the perimeter |
| Model | not disclosed on the service page | you choose it yourself and can replace it |
| Knowledge sources | project, ITS documentation, and BSP | your configurations, standard exports, policies, tasks, web search |
| Actions in code | agent mode with code and metadata editing, four confirmation modes | depend on the implementation; you design permissions and confirmations yourself |
| Response speed | in the Habr comparison — noticeably faster | in the Habr comparison — noticeably slower |
| Cost as of the check date | free until the date specified by the vendor; the future price has not been announced | development, infrastructure or API fees, maintenance |
| Who is responsible for quality | vendor | your team or contractor |
| Time to first result | install the plugin and enter the key | from prototype to production solution — depends on the volume of sources |
When to choose Naparnik, your own assistant, or both
Naparnik is enough if:
- the team already works in 1C:EDT and the tasks are routine: explain someone else's code, do a code review, find a bug, write comments;
- sending code to the vendor's servers is allowed under your security rules;
- you need to get started in the next few days, and you don't have the budget or people to build your own solution.
Your own assistant makes sense if:
- code, data, or questions cannot be sent outside the company perimeter, even to the platform vendor;
- a specific model matters, for example an open model running on your own servers;
- answers must rely on your policies, past customizations, and support history;
- the assistant is needed not only by developers, but also by consultants, users, or the support desk;
- the assistant must take actions: create tasks, search data in the database, prepare document drafts.
Both tools at the same time — a common practical option: developers use Naparnik in EDT for everyday code, while a custom assistant covers the knowledge base for customizations and user support. This way you don't build what the vendor already gave away for free, and you don't expect from it what it never promised.
When neither option is needed: if there are few customizations and they are one-off, and user questions can be handled with a short guide. In that case, start by cleaning up the documentation — an assistant won't help without sources anyway.
Data and security: what leaves the perimeter
Naparnik. According to the vendor, computations run on 1C company servers in Russia, configuration data is not shared with third parties, no external APIs are used, and the code is not used to train global models without explicit consent. At the same time, your project code still goes to the vendor's servers — this should be approved by your security team and, if the code contains personal data or trade secrets, by legal.
Custom assistant. The level of protection depends on the architecture. If the model is cloud-based, code snippets and questions are sent to the model provider; if the model runs on your own servers, the data stays within your perimeter, but hardware and maintenance costs increase. A breakdown of the options is in the articles about an AI agent for development in a closed environment and about the cost of cloud vs. on-prem LLM.
In both cases make sure the agent doesn't get more permissions than it needs: code changes only with approval, access to the production database read-only and only for the necessary objects, all changes through Git and code review. After heavy use of AI assistants, it's a good idea to separately verify code quality; more on that in the article about code auditing after vibe coding.
How to compare the options on your own tasks
Below is AIrassvet's editorial method for a pilot comparison. This is not the result of our own testing, but a process that lets you make a decision based on your own data.
- Collect 20–30 real tasks from the past few months: questions about the standard configuration, your customizations, errors from the log, and user requests. Don't take tasks whose answers are already known to the assistant from the prompt.
- Split the tasks into groups: "in the documentation," "only in your code," "neither here nor there." Results by group will tell you more than the overall score.
- Set evaluation criteria in advance: the answer is usable as is, usable with edits, incorrect, or dangerous. Separately note fabricated metadata objects and procedures.
- Have two developers evaluate each answer independently, without knowing which tool produced it. Review any differences in the evaluations together.
- Measure time to a usable result, not just response time: a fast but wrong answer costs more than a slower correct one.
- Calculate the total cost for your setup: development, infrastructure or API, maintenance, and index updates after each configuration release.
- Make the decision by task group: often the conclusion sounds like this — "daily code in Naparnik, questions about our customizations and user support in our own assistant."
Common mistakes
- Comparing on demo questions. On standard questions from the documentation, the gap between tools is smallest; questions about your customizations are usually what matter most.
- Building your own assistant on vector search alone. Exact names of objects and procedures get lost; you need hybrid search.
- Not updating the index. An assistant that knows a configuration from three months ago confidently recommends deleted procedures.
- Giving the agent write permissions without confirmation. The "no confirmations" mode is convenient, but code and metadata changes should go through review.
- Planning the budget based on the free period. The vendor's end dates vary, and the price after that has not been announced.
- Expecting the assistant to know things that are not in the sources. If procedures are not documented, neither Naparnik nor your own assistant will find them.
Frequently Asked Questions
What is 1C: Naparnik?
An AI assistant from 1C for developers, built into 1C:EDT. According to the vendor, it searches the project, explains and fixes code, edits metadata, reads EDT errors, works with Git, and searches the ITS and BSP documentation. According to 1C, processing is performed on its servers in Russia.
How much does 1C: Naparnik cost?
As of October 12, 2026, the service is free. The website code.1c.ai says “free until February 1, 2027,” while the ITS portal says no additional charge until December 1, 2026. The price after the free period has not been announced in either source.
Does 1C: Naparnik work in the Configurator?
In the official descriptions we checked, only 1C:EDT support is listed: the service is included in EDT 2025 and later, and for earlier versions it is installed as a plugin. Configurator support is not mentioned on these pages.
Can you build your own AI assistant for 1C using an open model?
Yes. The author of the comparison on Habr built an assistant using a LangGraph agent framework, hybrid PostgreSQL and pgvector search over exports from standard configurations, and web search. The model can run on your own servers, but that requires hardware and maintenance.
Is your own assistant better than Naparnik?
There is no clear answer. In the comparison across five questions, the self-built assistant gave more useful answers for some tasks, but it was noticeably slower, and the development costs were not included. The decision should be made based on a pilot using 20–30 of your own tasks.
Can both tools be used at the same time?
Yes, and this is a common practical option: Naparnik handles day-to-day coding work in EDT, while your own assistant handles questions about your customizations, procedures, and user support.
Primary sources and verification method
We reviewed the primary sources on October 12, 2026. The capabilities, requirements, security modes, and access terms for 1C: Naparnik are taken from official 1C pages; the stack and comparison results come from the Habr post. The pilot comparison method, selection criteria, and list of errors are AIrassvet editorial methodology, not the result of our own testing.
- 1C Company: 1C: Naparnik for Development — capabilities, security modes, data processing, free period until 02/01/2027.
- 1C Company, ITS portal: 1C: Naparnik for Developers — requirements, keys, free period until 12/01/2026.
- Yevgeny Sokolov, Habr, 10/10/2026: “1C Naparnik. Is It Worth It?” — the self-built assistant stack and a comparison across five questions.
Limitations: the Habr comparison was conducted by one author on five questions without numerical scores or independent verification; the vendor does not disclose the model used by Naparnik; the terms of free access conflict in two official sources and may change.
How AIrassvet helps with an AI assistant for 1C
If the vendor’s assistant is not enough — you need your own environment, your own model, or answers about your customizations and procedures — the assistant should be built around your knowledge sources and access rules. On the AIrassvet website the directions directly related to this task are described:
- RAG systems and enterprise knowledge bases — search across configuration exports, procedures, and customization history.
- Deployment and fine-tuning of language models, including on company servers, when code and questions cannot be sent to an external service.
- AI agents integrated with business systems, and agentic automation in the interfaces of accounting systems, including 1C.
- Process audit, data preparation, testing, launch, team training, and support.
The first step is to choose one group of tasks, such as questions about your customizations, describe how they are answered today, where the needed sources are stored, what data restrictions apply, and what criterion you will use to decide that the assistant’s answers are usable.
Bottom line: start with Naparnik, decide based on your own tasks
1C: Naparnik is the shortest path to an AI assistant for 1C developers: it is built into EDT, runs on the vendor’s servers, and is free at the time of review. A custom assistant is needed when the environment, model choice, and knowledge the vendor does not have matter: your customizations, procedures, task history, and user questions. The Habr comparison shows that an open stack with strong search can match the vendor solution in answer quality, but pays for it with speed and development effort.
Do not choose based on reviews. Assemble 20–30 of your own tasks, split them by source of the answer, have two developers independently rate both options blindly, and calculate the total cost. Most often the conclusion is not “either/or,” but “both”: Naparnik for everyday code and your own assistant where you need your context.