Integrations
How to connect an AI agent to an ERP over MCP
Akeda ERP works as an MCP server. A compatible client connects to a single address, https://erp.akeda.ru/mcp, the employee signs in with their own account, and the agent can reach more than 1,700 operations: stock balances, document posting, financial reports, tasks, chats and the knowledge base. By default the client sees seven tools and finds the operation it needs by search. The agent inherits that employee's role: what the person cannot see, the agent cannot see either. No separate key, no local wrapper, no database export.
| Server address | https://erp.akeda.ru/mcp |
|---|---|
| Operations | more than 1,700; by default the client sees 7 tools: search, call, script and service ones |
| Protocol | Streamable HTTP MCP, OAuth 2.1 with PKCE and dynamic client registration |
| Clients | Claude, OpenAI Codex and other MCP-compatible clients |
| Permissions | the employee's role in the workspace; agents have no separate permission list |
| Workspace | chosen inside the session, not fixed in the connection |
| Access token | lives one hour, then refreshes |
| Machine access | an API key scoped to one workspace, read-only by default |
AI agent and AI assistant for business: the difference
The words “AI assistant” and “AI agent” are used interchangeably, though the practical difference matters. An assistant answers a question: you ask, it tells. An agent carries out a task of several steps — it decides where to look for the data, cross-checks what it finds and returns a finished answer rather than a hint.
For a company that means this. An assistant has to be told where to look. An agent only needs the task: “pull everything on this customer — orders, shipments, payments, outstanding debt, open tasks” — and it walks the sections itself.
The difference between “an AI agent” and an agent that is actually useful comes down to one thing: whether it can reach your data. An agent without access retells the internet. An agent with access answers with a figure from your books. Everything below is about how that access is granted and what limits it.
What MCP is and why an ERP needs it
MCP is an open protocol an AI assistant uses to reach an external system: to ask for data and to perform actions. For an ERP it means the assistant stops being a chat in a separate window and starts working with your real balances, payments and tasks.
The usual route to that result is a custom integration per assistant, plus exports and the constant worry that the copy has drifted from the original. MCP removes that layer: the system declares its tools once, and any compatible client sees them.
What the agent does in Akeda
Tools mirror the sections of the system. The main modules as of October 2026 are below; the server exposes more than 1,700 operations in total, including service operations for the platform owner.
| Section | Operations | What the agent can reach |
|---|---|---|
| Directories and documents | 300 | contacts, products, employees, documents, posting, registers and reports |
| Finance | 272 | accounts, operations, statements, payment calendar, P&L, cash flow, settlements |
| Stock | 181 | balances, reservations, batches, receipts, shipments, inventory counts, purchasing |
| Tasks and calendar | 173 | projects, sections, tasks, checklists, sprints, meetings and booking slots |
| CRM | 171 | customers, deals, requests, sales plan |
| Document flow and e-documents | 134 | deal paperwork, approvals, payment requests, exchange through e-document operators |
| Marketplaces | 105 | stores, catalogue, supplies, FBS picking, finance weeks and platform reports |
| Product marking | 99 | Chestny Znak codes, code orders, boxes, labels, GIS MT documents |
| Knowledge base and files | 61 | spaces, pages, search and answers with citations, drives, folders and sharing |
| 53 | mailboxes, messages, drafts, rules, attachments | |
| Chats | 44 | conversations, messages, mentions |
| Settings | 29 | legal entities, roles, members, custom fields, API keys |
The list is not written for a brochure: it is generated from the sources during the build, and the code map verifies that every tool has a live API route behind it.
Compact mode: seven tools instead of a thousand
The full list of more than 1,700 operations would take about a megabyte of model context before the first question. So by default the server shows a client seven tools: the agent rules, the list and choice of workspace, the catalogue, operation search, an operation call and a short read-only script.
The agent searches for an operation in plain words — “invoice a customer”, “overdue tasks” — and runs it with akeda_go_call. To combine data from several sections at once it writes a short function for akeda_go_execute, which reads, joins and returns only the result. Changes never go through a script: each is a separate call, and when a confirmation is required the agent shows the person a summary and waits for an explicit yes.
Every such call is checked exactly like a direct one: the employee's permissions, the modules switched on, the confirmation. A client that needs the whole list at once can use https://erp.akeda.ru/mcp?toolset=full.
Connecting in three steps
- In your client, add the MCP server at
https://erp.akeda.ru/mcp. Client registration is automatic — there is no application id to paste. - Sign in to Akeda with your own account. Only the Akeda sign-in page accepts the password; the client never sees it.
- Pick the workspace. With several workspaces the session starts without a default, so an agent cannot quietly act in the wrong company.
From there the agent sees the tools and can, for example, build a stock report, find a document, create a task or record a receipt.
Permissions: an agent can never exceed the person
There is no separate permission table for agents, and that is deliberate. The agent acts as the employee who signed in: a section closed to them is closed to the agent, and a module switched off in the workspace publishes no tools at all.
The access token lives an hour and refreshes; the connection is revoked where it was granted. Every call is checked on the API side, not in the client.
Machine access with an API key
When the caller is your script or service rather than a person with an assistant, use an API key. The key is scoped to one workspace, grants read-only access by default, is shown once at creation and stored only as a hash. It can be revoked, it expires, and it is audited.
The split is intentional: user access and machine access never mix, so revoking an integration key does not sign people out.
How this differs from AI built into the product
A built-in assistant can do what its author implemented. An MCP agent brings its own head: you choose the model and the client, and the system supplies the tools and the boundaries. Switch assistants and there is no integration to rewrite.
The honest flip side: answer quality depends on the model you picked, not on Akeda. The system is responsible for the data being correct and the action being permitted.
What the agent will not do
- The agent does not bypass permissions: if a section is closed to the employee, the tool returns a refusal.
- A module switched off in the workspace exposes no operations — they are absent from both search and the list.
- Document posting follows the same rules as the interface: an invalid document is posted neither by a person nor by an agent.
- The password never reaches the client: sign-in happens on Akeda's side.
- The set of operations changes with the product, so the agent finds them by search while working instead of caching a copy of the list.
Frequently asked questions
- Which ERP has an MCP server?
- Akeda ERP. The server runs at https://erp.akeda.ru/mcp and exposes more than 1,700 operations across the sections of the system: stock, finance, CRM, tasks, document flow, marketplaces, chats and the knowledge base.
- How do I connect Claude or ChatGPT to a company's ERP?
- Add the Akeda MCP server at https://erp.akeda.ru/mcp in your client, sign in with your Akeda account and choose the workspace. No separate key or local wrapper is needed, and client registration is automatic.
- Will the agent see all company data?
- No. The agent acts with the permissions of the employee who signed in: a section closed to the person is closed to the agent, and a disabled module publishes no tools.
- Can a script get access instead of an assistant?
- Yes, through an API key. It is scoped to one workspace, read-only by default, shown once, stored as a hash and revocable at any time.
- Do I have to export the database for the agent to work with it?
- No. The agent reaches live data through tools; no copy of the database is created or kept on the client side.
- What happens if we switch assistants?
- Nothing needs rewriting: any MCP-compatible client connects to the same address and receives the same list of tools.
- What is an AI agent for business in plain words?
- A neural network given access to a company's working systems and a task of several steps. Unlike an assistant, which answers a single question, an agent walks the sections itself, cross-checks the data and returns a finished answer. Only an agent with access to your own data is worth having.
- Does a small company need an AI agent?
- It pays off where answering “how are we doing” means opening three or four sections and adding the figures up by hand. If the whole business fits in one spreadsheet, an agent adds nothing. If sales, stock and money already live in one system, connecting one takes minutes and needs no implementation project.
See Akeda
with your workflow.
Tell us what you want to connect or automate. We will show how the process could work in Akeda, what data it needs and where to start.
- 01Understand your needs
- 02Explore connected workflows
- 03Define the next step
Request received
Thank you! We have your number and will get in touch on the next working day.
What happens next: a short conversation about your process — what to connect, which data is needed and where to start. No sales pitch.