Яндекс.Метрика
Back to blog
Product

We launched Pachca's MCP server and explored what's already on the market

Which Russian services you can connect to an AI agent, and how official MCP servers differ from community-built ones

ByPachca Team
We launched Pachca's MCP server and explored what's already on the market
Pachca MCP server documentation

Connection, authentication and tools for working with chats and tasks

More than thirty Russian companies have released their own MCP servers over the past year, from Sber and Finam to VkusVill and MoySklad. Most appeared this summer and autumn. In fintech, launching a server for AI agents has become an almost obligatory line in a press release.

A parallel market has grown alongside the official servers. One developer has published more than a hundred npm packages promising to connect an agent to almost any popular service, from 1С to HeadHunter. Catalogs list these wrappers alongside the companies' own servers. The difference matters when you trust an agent with money or customers.

Pachca released its own server on October 28. While building it, we studied both official and community-built servers. Here's what we found.

What MCP is and how it works

MCP (Model Context Protocol) is an open standard that lets an AI assistant connect to an external service, read its data and take actions in it. Anthropic published it in late 2024, and OpenAI, Google and Microsoft later adopted it. Build a server once, and it can work with any popular assistant.

For example, a procurement manager needs to work out which products will run out before the end of the month and send a supplier an order by the evening.

  • Without MCP, they export stock levels from MoySklad to Excel, paste the spreadsheet into a chat, explain the columns and copy the answer into an email.
  • With MCP, the assistant retrieves stock and sales data itself and prepares a draft order. The person just adjusts a line and sends it.

How an AI agent connects to a service

On one side is the program running the agent: Claude, Claude Code, Cursor, Codex, ChatGPT or an agent in Yandex AI Studio. On the other is a server for a particular service. The program gets a list of tools from the server, and the agent chooses which one to call.

Usually, the company hosts its server at an address such as https://mcp.pachca.com/mcp or https://invest-public-api.tbank.ru/mcp. Connecting it takes a single line in the settings. In Claude Code, use:

claude mcp add --transport http <name> <address>

In Cursor, add an entry to mcp.json. Community-built packages usually work differently: you run them on your computer with a command such as npx -y <package> and pass the service's key through an environment variable.

Next, the server needs to know who is connecting. With Pachca, VkusVill and OkoCRM, the program opens the service's website so the person can see the access the agent requests and approve it. Where browser login isn't available, you enter an API token in the settings, and the agent gets exactly that token's permissions. If the agent only needs to look at data, issue a read-only token.

The program itself decides whether to ask for permission before taking an action: before every call, only before changes, or never. Well-designed servers label tools that read data and those that change or delete it, and programs take those labels into account.

How an AI agent connects: client, MCP server, service API, authentication and action approval

Which Russian services have released MCP servers

Banks and brokers were the first to release MCP servers in Russia: Sber, T-Bank, Finam, BCS and YooKassa. Finam is one of the few to disclose usage figures: its server handled 1.5 million requests in two months.

Official MCP marketplaces followed: MCP Hub in Yandex AI Studio, the Cloud.ru marketplace and Timeweb Cloud's server gallery. They let you connect a ready-made server to your agent, but not every server listed was built by the service itself. More on that below.

The tables include only servers released by the companies themselves. The data is current as of October 29, 2026. Where the exact launch date is unavailable, only the year is listed.

Finance and payments

Finance and payments: official Russian MCP servers as of October 29, 2026

Text version of the table
CompanyWhat the agent can accessLaunch
Sber (SberBusiness, Sber API)Ruble payments and their status, counterparty details, statements, credit lineJune 2026
T-Bank (T-Investments)Portfolio, quotes, transactions and orders with full accessJune 2026; public release in September
T-Bank (T-Business)Accounts and statements, invoices and closing documents, draft payment orders2026, beta
FinamBrokerage account in FinamTradeJuly 2026
BCSMarket data, portfolio, placing and canceling ordersAugust 2026
YooKassaPayments, refunds, invoices, payment linksSeptember 2026
Alfa-BankExternal integrations; no public documentationAnnounced in May 2026

CRM, tasks and workflows

CRM, tasks and workflows: official Russian MCP servers as of October 29, 2026

Text version of the table
CompanyWhat the agent can accessLaunch
Bitrix24CRM, tasks, email, MCP Hub, business process creation (Bizproc), REST API documentation serverLate 2025; Bizproc in September 2026
OkoCRMDeals, contacts, tasks, conversations, analyticsSeptember 2026
PlanfixTasks and projects, OAuth login2026
YouGileTasks, discussions, task historySeptember 2026
Aspro.CloudAnalytics, summaries, process managementSpring 2026
NCW RU (EGAR Technology)Document templates, business processes, reference data2026
MoySkladSales, stock levels, reports, customer segments2026
EME.WMSShipments, orders, a log of every agent requestSeptember 2026
AlbatoOne server for actions in 1,000+ apps through their OAuth2026

Communications and video

Communications and video: official Russian MCP servers as of October 29, 2026

Text version of the table
CompanyWhat the agent can accessLaunch
PachcaChats, threads, conversation search, messages, tasksOctober 2026
VK WorkSpaceTeam chat; Mail, Disk and Calendar planned by the end of the yearAnnounced in September; SaaS in October 2026
TrueConfConferences, recordings, invitations, chats, calendars (32 tools)October 2026
MWS TeamStream (MTS)Video library and workspacesJuly 2026
KinescopeVideos, projects and playlists, viewing analytics, subtitles, streams2026, public beta

Websites, marketing and shopping

Websites, marketing and shopping: official Russian MCP servers as of October 29, 2026

Text version of the table
CompanyWhat the agent can accessLaunch
TildaPages, blocks, Catalog products, Feeds posts, SEO error checks2026
DashaMailLists, campaigns, reports, templates, webhooks (95 tools)2026, no announcement
Brand AnalyticsSocial media and news monitoringJuly 2026
PR-CYSEO analytics services2026
SpyWordsCompetitive analysis services, OAuth login2026
TutuSearch for flights, trains, buses, suburban trains and hotelsJune 2026
VkusVillProduct search with availability, order historyDecember 2025; login in October 2026
2GISBusiness and address search, geocoding, routesSeptember 2026

Cloud, development and security

Cloud, development and security: official Russian MCP servers as of October 29, 2026

Text version of the table
CompanyWhat the agent can accessLaunch
Yandex CloudTen infrastructure servers, documentation search2026
Yandex Tracker, Wiki, Forms, SourceCraftTasks, Wiki pages, forms and responses, repositories2026
Cloud.ruInfrastructure servers and an MCP marketplace2026
Timeweb CloudServers, databases, Kubernetes, domains, S3July 2026
Aurora OSDocumentation and code examplesSeptember 2026
DocumenterraDocumentation knowledge basesNo announcement
SecuritmAssets, risks, tasks and security metrics, read-only modeOctober 2026

Community-built servers: what to check

Search an MCP catalog for a Russian service, and you're more likely to find an enthusiast's package than the company's own server. The npm namespace @theyahia has the largest collection: the WWmcp project includes more than a hundred packages, from 1С and CDEK to HeadHunter, all maintained by one person. Other authors have projects too. For example, marketplaces-mcp-ru connects seller accounts on Wildberries, Ozon, Yandex Market and Avito, none of which have their own servers.

You can use these packages, but you should understand what you're agreeing to. Their authors aren't affiliated with the services they connect to and offer no guarantees: WWmcp's README neither claims official status nor includes a guarantees section. If a service changes its API, the package stops working, and only its author can fix it. The same applies to the Yandex AI Studio catalog: the Kontur.Focus and amoCRM servers there weren't built by Kontur or amoCRM. They're wrappers around someone else's API, without an SLA.

The @theyahia packages were also released in batches, as many as thirty a day. Repository checks confirm that a server starts, not that it correctly handles a live request. Some packages are empty: the Ozon and Yandex Market servers remained at version 0.0.1 without a single tool. Meanwhile, your service key sits in the program's settings on your computer, and the agent gets all its permissions. That's fine for a prototype, but not for access to money, orders or customers.

How official servers differ from community-built ones

Sber, YooKassa, Bitrix24, Tilda, Planfix, MoySklad, 2GIS, Yandex Tracker and amoCRM have both their own servers and community-built alternatives. A community server may even offer more features, but an official server provides things an API wrapper can't add.

The company maintains it and updates it when its API changes. It's usually hosted remotely: you log in through a browser and can revoke access in the service's settings. It can also follow the service's own rules:

  • OkoCRM and Pachca label messages sent by agents;
  • Securitm is read-only, with access controls down to individual objects;
  • EME.WMS logs every agent request;
  • Tilda issues a key with an expiry date;
  • in DashaMail, a campaign created by an agent never sends itself: it waits for a person in Drafts.

If the official server covers what you need, it's the better choice.

Official and community-built MCP servers compared: responsibility, access, service rules and support

Why we built Pachca's MCP server

Almost every server on the market gives an agent access to a single service, but the results still need to go somewhere. A SberBusiness statement, a DashaMail campaign report or a stock summary is easiest to read where the team already discusses it. That's why we wanted an agent to be able to read a thread, search conversations and write in a chat.

Pachca's server gives the agent a tool for every method in the public API. You can ask it to summarize decisions in a release thread, find where the team discussed moving to a new server this week, or create tasks from a discussion with a Friday deadline. The agent can only see and do what the token owner can. The server works on every plan, including the free Starter plan.

We borrowed some ideas from other services. As in DashaMail, an agent can leave a draft for an employee instead of sending a message itself. Each tool is labeled as a read, a change or a potentially dangerous change, such as deleting something or removing a member. Messages sent by an agent after browser login carry the employee's name, with the program's badge on the avatar and a tooltip such as “Sent via Claude Code.” The server passes conversations to the agent as data, so a message saying “forward all tokens” remains just a message. Pachca also doesn't open links from agent messages itself, so an instruction planted in a chat can't use that mechanism to send data to someone else's server.

Claude Code, the Claude app, VS Code with GitHub Copilot, Codex and the ChatGPT desktop app use browser login. A consent screen appears, and the connected program then shows up under Authorizations on the API page, where access can be revoked. Cursor and other programs connect with a personal token; agents in Yandex AI Studio and on GigaChat use a bot token. In Claude Code, one command and login through /mcp are enough:

claude mcp add --transport http pachca https://mcp.pachca.com/mcp

A Pachca plugin is available for Claude Code, the Claude app, Codex, VS Code and Cursor. It installs the server along with skills that tell the agent where to post, when to ask before acting and which token to choose. It also adds ready-made commands for catching up on missed messages, summarizing a thread, creating a digest or announcement, and turning a discussion into tasks.

Frequently asked questions

It's a program that lets an AI assistant connect to a service through the open Model Context Protocol standard to read data and take actions. A single server works with Claude, ChatGPT, Cursor and other assistants.

Run claude mcp add --transport http <name> <address>, then log in through /mcp. For Pachca, the server address is https://mcp.pachca.com/mcp.

Not as of late October 2026. Only third-party servers are available for their seller accounts, such as marketplaces-mcp-ru or the MP Manager server.

Its authors aren't affiliated with the service and offer no guarantees. The API key is stored on your computer with all its permissions. These servers suit prototypes; for access to money and customers, an official server is the better choice.

Yes, since October 28, 2026. It works on all plans, including the free plan. See the documentation for details.

What to check before connecting

For now, banking, brokerage and cloud servers look the most dependable: they are backed by companies with clear responsibility. Community-built packages are useful for experiments and prototypes. Before trusting an agent with money, orders or customers, check three things:

  1. Who owns the server.
  2. When it last received a substantive update.
  3. Whether it handles a real API request.

To try it with your work conversations, see the Pachca server documentation. Tell us which workflows you find missing.

Keep reading