What should an MCP server for your business processes expose, and what should it refuse?
Expose the shape of the work: each process as steps and flow, who is responsible and who is accountable, the applications and data each step touches, the requirements, decisions and controls that apply, and your organisation's own definitions, each with a permanent identifier so the assistant cites rather than guesses. Refuse every financial figure, personal contact details and raw HR records, anything the person asking has not been granted, and any tool that writes a record directly: a change should arrive as a proposal a person accepts.
Written for technology and operations leaders being asked to connect Claude, ChatGPT, Cursor or an internal agent to the organisation's process knowledge. Reviewed by Mike Bentley, founder, .
What MCP is, in the project's own words
The Model Context Protocol project describes MCP as "an open-source standard for connecting AI applications to external systems". In practice an MCP server is a small service that an assistant such as Claude, ChatGPT or Cursor connects to, asks questions of, and may ask to act. Anthropic's help centre describes MCP as an open standard created by Anthropic, and the project's own site lists Claude, ChatGPT, Visual Studio Code and Cursor among the applications that support it, as at 9 November 2026.
According to the project's versioning page, the current version of the specification is dated 28 July 2026. It says a server can offer three things: resources, defined as context and data for the user or the AI model to use; prompts, templated messages and workflows for users; and tools, functions for the AI model to execute. Tools carry the risk, because a tool is something the model decides to call.
The specification is direct about that. It says tools represent arbitrary code execution and must be treated with appropriate caution, and that hosts must obtain explicit user consent before invoking any tool. Anthropic's help centre tells people to review Claude's tool approval requests and to choose "Allow always" only for servers and tools they trust, and OpenAI's developer documentation says ChatGPT requires manual confirmation in a conversation before a write action is taken. Those are good defaults, not a substitute for a server designed to refuse.
Why a process model is worth exposing
Ask an assistant with no connection to your organisation who approves a refund above a delegated limit and it answers from what organisations in general do. The answer reads well and may be wrong. Connect it to a model of how your organisation actually works, who does what, which application each step runs in, which controls apply and which requirements have been committed, and the assistant can answer from your record, name the step it read, and say when that step was last reviewed.
A process model is a structured record of how an organisation's work happens: each step, who does it and who answers for it, the application it runs in, and the requirements and controls that apply. It is a better thing to expose than a document store. Documents disagree with each other, and an assistant reading three versions of a procedure picks one without telling you. A model holds one answer per fact, with an owner and a date, and a conflict shows up as a conflict rather than disappearing into a fluent paragraph.
Much of a process model is also already written down somewhere: on the wiki, in the onboarding deck, in the last audit pack. What changes when an assistant can read it is not confidentiality but reach. More people get a grounded answer, and fewer of them have to find the one person who knows.
What to expose
An MCP server for business processes should expose the facts an assistant needs to describe how work happens, each with a permanent identifier so an answer can cite the record rather than paraphrase it. Six kinds of record cover it.
Every answer should carry its provenance: the identifier of the record, who owns it, and the date it was last reviewed. An assistant that can say which step it read, who owns that step and when it was reviewed is useful. One that says procurement usually handles this is not, even when it happens to be right.
- Process maps as steps and flow: what happens, in what order, with the branches and the hand-offs between teams.
- Who is responsible for doing each step, and who is accountable for it: one named position for the second, never a team.
- The applications and data each step touches, so the assistant can say where a step is done and what it reads and writes.
- Requirements and decisions, each with an identifier that never changes, so a reference made today still resolves a year from now.
- Controls and the policies they implement, mapped to the steps and requirements they cover, and the gaps where no control is recorded.
- Your organisation's own definitions of its own terms, so the assistant uses your words rather than the industry's.
What to refuse, and why
Four categories should never come back from an MCP server's tool call, even where the person asking could read them on a screen inside the system that holds them.
Each refusal has to live in the server's code, not in a label. The specification lets a server mark a tool read-only, and Anthropic's connector checklist says such tools can run in Claude without per-call confirmation while destructive tools always prompt. But the specification also says clients must treat tool annotations as untrusted unless they come from a trusted server, and OpenAI's documentation warns that a write can occur even when a server marks an action read-only.
- Pay rates, prices, budgets and any other financial figure. A model should never generate or reason over money it cannot verify: it produces a confident number, the reader cannot tell a retrieved figure from an invented one, and the figure travels. Names, counts and statuses can be narrated; amounts stay where they are owned.
- Personal contact details and raw HR records. A step needs an accountable position, not a phone number or a personnel file.
- Anything the person asking has not been granted. The server answers as that person, and someone with no grant sees nothing rather than a redacted version of everything.
- The ability to write a record directly. Every change should arrive as a proposal a person reviews and accepts row by row, so the record stays one a person stands behind.
How to scope the connection
Connecting an assistant to a process model is an access decision. Anthropic's help centre adds three details worth knowing: on Team and Enterprise plans only designated owners and roles can add a custom connector; a connector set up with a fixed credential is shared by everyone who uses it; and in Claude's research mode tools can run without further approval, so it advises disabling write tools there. Prefer a connection where each person signs in as themselves.
Four rules keep the connection narrow.
- One engagement or body of work per credential. A token that can read everything will eventually read everything; bind each to the one body of work it serves.
- Grants that default to no access. The specification allows the list of tools a server returns to vary with the access presented, so a person with no grant should see no tools rather than a polite empty answer.
- Tokens that expire and can be revoked. The specification's security guidance says a server must not accept a token that was not issued for it, and warns against broad scopes in favour of a minimal starting set. You need a way to cut a connection off without waiting for expiry.
- An audit record of what was read. The specification tells clients they should log tool usage for audit; your server should keep its own record of every call, with the identity, the grant and the time, because the client's log is not yours.
Six questions to ask a vendor
A demo shows the assistant answering well. These six questions find out what happens when it should not.
The sixth question, on tool descriptions, comes from Anthropic's pre-submission checklist for connectors in its directory, which rejects a single tool that mixes reads and writes, and rejects descriptions that tell the model how to behave, pull instructions from elsewhere or carry hidden text. Both Anthropic and OpenAI say in their documentation that a custom server is not verified by them. The checking is yours.
- Which fields does the server refuse to return, and is the refusal enforced in the server's code or only described in a tool's text? Ask to see a refused call.
- Does any tool write a record directly? The answer you want is that every change arrives as a proposal a person accepts, and that you can see one.
- Is a credential bound to one engagement or body of work, and what does a caller with no grant see: no tools, an error that says so, or a thinner version of everything?
- Does the server check that a token was issued for it and nothing else, and can we revoke a token from our side today?
- Is there a log of every call, with who asked, under which grant, when, and what came back, and can we export it ourselves?
- Are read and write tools separate and annotated, and can we read every tool description before we connect, to check it describes the tool rather than instructing the model?
Three prompts, and what a good answer looks like
Who is accountable for approving a new supplier, and where? A good server returns the step, the accountable position, the application the approval is recorded in, the step's identifier and the date of its last review, and the assistant cites each.
Which controls apply to the refund process, and which requirement does each satisfy? The server returns the controls mapped to the steps they cover and the requirement each is recorded against, and names the steps with none recorded. A good answer includes the gap; one that reads as complete when the record is thin is the failure to watch for.
What does the finance team get paid, and what did the refund process cost us? The server refuses. A good refusal says what it cannot do and where the work is done by hand: this connection does not return pay or cost figures; they are held in the finance system and reported by finance. It never fills the gap with an estimate or pretends the question was unclear.
Omni by Valstra's MCP server is one built to these rules: the records (steps, owners, applications, controls, requirements) are readable within a grant, financial figures are withheld from the model, and every change arrives as a proposal a person accepts row by row; the product page and the page on how Valstra constrains the AI, linked below, set out the detail. Whatever you connect, run the third prompt and read how it says no.
Sources
- Model Context Protocol, What is the Model Context Protocol?
- Model Context Protocol, Specification (current version 2026-07-28)
- Model Context Protocol, Versioning
- Model Context Protocol, Specification 2026-07-28, Server features: Tools
- Model Context Protocol, Security Best Practices
- Claude support, Get started with custom connectors using remote MCP
- Claude Docs, Connector pre-submission checklist
- OpenAI Developers, Building MCP servers for plugins and API integrations
Common questions
- Is MCP only for Claude?
- No. The project describes MCP as an open-source standard, Anthropic's help centre says Anthropic created it, and the project's site lists Claude, ChatGPT, Visual Studio Code and Cursor among the applications that support it. A server built to the specification can be reached from any of them, and from an internal agent you build yourself.
- Does connecting an assistant mean it can change our records?
- Only if the server offers a tool that writes. The specification says hosts must obtain explicit user consent before invoking any tool, and both Anthropic's and OpenAI's documentation describe confirmation before a write, but the durable protection is a server that exposes no direct write and turns every change into a proposal a person accepts.
- What stops the assistant quoting our salary data?
- Nothing in the protocol: MCP carries whatever the server returns. The server has to refuse the field, so pay, prices and every other figure should be absent from every tool's output. Enforce that in the server's code, not in a tool's description, because the specification tells clients to treat a tool's annotations as untrusted unless they come from a trusted server.
- Can we see what the assistant asked?
- Yes, if the server keeps a log. The specification tells clients they should log tool usage for audit, and a well-run server keeps its own record of each call with the identity, the grant and the time; ask a vendor to show you the log and how you export it.
See it against your own processes
A demo runs on your material, not a canned dataset. Bring a process you find hard to explain.