In plain words
Configured, coded, or connected
The core of the platform is never rebuilt for a business. What a business needs is added around it, in one of three ways. The third is for the systems it already runs.
The typical case
ConfiguredA playbook, no code
- Fields to collect
- Outcomes and their criteria
- Handover to a person
A case of its own
CodedAn app, inside the core or as a service of its own
- A step of a business process
- A new tool, such as ads analytics
A system the business already runs
ConnectedIts own tools, over MCP
- ERP: stock, orders, quotes
- Shop, ticket desk, booking system
What a conversation looks like
- Once, when it is connected
- Assistant to Your system: Which tools do you have?
- Your system to Assistant: 7 tools: find_parts, get_stock, get_order … · 6 only read
- A question
- Visitor to Assistant: Is the spindle bearing set for the FZ 650 in stock? We’d need two.
- Assistant to Your system: get_stock {part_no: FZ650-SB-01}
- Your system to Assistant: 4 on hand, 1 reserved, 6 arriving
- Assistant to Visitor: Yes: 3 available, two sets ship from stock.
- A change
- Visitor to Assistant: Please request a quote for the two sets.
- Assistant to Visitor: Confirm: request a quote · C-1042 · 2 × FZ650-SB-01
- Visitor to Assistant: Confirm ✓
- Assistant to Your system: create_quote {customer_no: C-1042, lines: …}
- Your system to Assistant: Q-26-2041 · € 2,960.00
- What does it need from us?
- An MCP server in front of the system. MCP, the Model Context Protocol, is an open standard, and more and more products ship a server of their own. For the others it is a thin layer over the API the system already has. The assistant needs its address and, if the server asks for one, a key.
- Can it change our data on its own?
- No. A tool that reads runs when the assistant needs it. A tool that writes runs only after the visitor presses Confirm, under a message that lists exactly what will be sent. And only if writing was allowed for this assistant at all.
- What does our system learn?
- What a call carries: the values the assistant filled in, and, if switched on, a number for the conversation. Not the visitor's contact details, not the conversation itself.
- Which tools does it get?
- The ones chosen for it from the system's list. The list is shown in full before the connection is switched on.
Under the hood
The tool list is read once, not on every turn
The platform speaks MCP revision 2026-07-28 over Streamable HTTP with the official SDK, version 2. A server of the 2025 revisions is served as well: the client negotiates the version and remembers it.
- The catalogue lives in the database. It is read when the connection is checked and refreshed in the background when the server says it is stale. A turn of the assistant only reads the stored copy, so a slow server never slows down the first word of an answer.
- Every schema is checked before the model sees it. A tool whose input the model provider would reject is left out with its reason, and the other tools stay. Names become
mcp__erp__get_stock, sorted, so the prompt cache does not depend on the server's order. - "Only reads" has to be said. A tool counts as reading when the server marks it so. Anything else is treated as writing: the safe reading wins.
A change is confirmed by the visitor, not by the model
When the model calls a writing tool, the call is not sent. The platform keeps it for 15 minutes and shows the visitor a message with Confirm and Cancel. The text of that message is built by the platform from the arguments, not written by the model, so the visitor confirms exactly what the server will get.
- One press, one call. The waiting call is claimed in one step, so a double click runs it once. If the tool list or the permission changed in the meantime, nothing is sent and the visitor is told so.
- One deadline for connecting and calling. When the model has been told that the system did not answer, a late write can no longer arrive behind its back.
- Every call is in the trace of its turn: server, tool, protocol, time, whether it was confirmed.
The address is the company's, the call is the platform's
- No way into the internal network. Only https. A name that resolves to a private address, the cloud's metadata service included, is refused at the moment of connecting, not only when the address is saved.
- One company's fault stays its own. A server that keeps failing is paused for half a minute, per assistant and address, so a wrong key of one company does not pause the same server for another.
- What comes back is data, not instructions. Tool descriptions and results are treated like passages of the knowledge base: read, never obeyed.
The demo ERP is a real MCP server
The parts & orders demo runs on a made-up ERP: 26 parts, three customers, ten orders. It is an MCP server of its own, and the assistant reaches it the same way it would reach a real one. Each conversation gets its own copy of the data, so the page next to the chat can show every call and lets you change a stock level or an order status.
Any MCP client can use it, not only the assistant. In Claude Code:
claude mcp add --transport http tannried-erp https://mcp.runkoworks.com/erpIn Claude Desktop or claude.ai it is a custom connector with the same address. Such a client first calls open_sandbox and gets a copy of the data of its own.
Connected works both ways
The assistant's own knowledge, the four documents it answers from, is an MCP server as well: every assistant on the platform has an address of its own. It offers one tool, search_knowledge, and it only reads. No model runs on the platform for it: your client does the thinking, the platform finds the passages and names the documents they come from.
claude mcp add --transport http tannried-knowledge https://mcp.runkoworks.com/bots/8e622427-8b5e-4458-8be0-3c875c349180With both added, Claude Code answers the questions of the demo the way the assistant on the page does: the terms from the documents, the numbers from the ERP. The same address works as a custom connector.
See it work: parts & orders on a live ERP. The rest of the platform: How the assistants are built.
The platform under these demos is my own. It runs live, and it is not open to external clients yet.
As of October 2026.