What is MCP, in plain words?
The Model Context Protocol, MCP, is a standard way for an AI assistant such as Claude to call the tools an application offers, for example "list the signals of this project". The application runs the tool and returns the data, and the assistant turns that data into an answer. MCP is the connection, not the intelligence: what Claude can say depends on what the tools return.
Two things are easy to confuse. The connector is a way to ask questions and file proposals. It is not a second copy of the database, and it is not an assistant that acts on its own. A question returns what the system holds at that moment, and a proposal does nothing until a person approves it.
For a construction project this matters because the data lives in several places. A connector lets a question be answered from the system itself, rather than from a copy someone exported last week.
What can Claude read in Reperline?
The Reperline connector offers read tools for each project a person can see. They return the project status, the milestone schedule and its variance against the plan, the budget plan against the actual, the cash flow, the inventory of units with their sale status, the tenders and their bid comparisons, the risks and signals, and the control rules and thresholds in force.
Document details are readable too: the type, title, language and extraction status of each file. The extracted data of a document is returned only when a person has confirmed it. Anything else is marked as unconfirmed and must not be presented as a budget fact. The file itself is not served through the connector.
Answers about money come from the database, that is, from confirmed data, not from reading a file in a chat.
What can Claude change, and what can it only propose?
Claude cannot change a record. It can file four kinds of proposal, and each one waits in the Inbox until a person decides:
- A milestone progress update.
- A new budget line.
- A transaction read from an invoice, payment, commitment or credit note, with the document name and link kept on the proposal.
- A risk the rule engine cannot see, such as rumours about a supplier. Rule based signals are calculated by the system, so the connector does not log them as risks.
How does a proposal become a record?
A proposal is only a proposal. A person with the right role opens the Inbox, checks it against the source, and approves or rejects it. Transactions can be proposed only by the owner or finance role, so other connected users cannot file them. Once a proposal is approved, the change and the connection that filed it are both recorded in the activity log.
Treat the Inbox as a check, not a formality. The proposal keeps the document name and its link, so comparing the figures with the document takes a moment. An approval given without that comparison is the one place where a wrong figure can enter the budget.
The system also refuses a duplicate. A transaction with the same project, kind, counterparty and document number cannot be filed twice, so an invoice that was sent to the connector twice does not enter the budget twice. Where Claude is unsure of a figure, it should ask the user rather than guess.
Whose rights does Claude use?
Each person connects Claude under their own account, so the connector can do what that person can do in the web app and nothing more. The database enforces this through row level security, so a person who cannot see the budget will not get budget data through Claude either. Anyone who uses Claude needs an account with a role, set by the owner.
A practical rule follows. If a person cannot see a budget line in the web app, they should not expect Claude to show it to them in a chat. The fix is to ask the owner to change the role, not to rephrase the question.
The remote connector uses OAuth. In Claude Desktop or claude.ai, a team administrator adds the Reperline connector using the address of the app. Each user then presses Connect, signs in to Reperline and allows access. Ask your administrator how the connector is set up in your Claude workspace.
What questions can the office ask Claude about a project?
The most useful questions are the ones whose answer changes what someone does next. A few examples show the range. The answer comes from the tools at the time of the question, so it is only as current as the last fact entered, and Claude should say so when data is missing.
Claude also cannot see what has never been entered: a phone call with a subcontractor, a verbal change agreed on site, or a delivery that has not been recorded. Those stay invisible to every answer until someone enters them, which is the same limit the office has.
- "Which milestones are forecast to end after their planned date, and by how many days?"
- "Which budget items have a forecast above plan, and which commitments moved them?"
- "What does the bid comparison of the open tender show, and how far is each bid from the lowest?"
- "Which buyer instalments are overdue, and what is the cash position this month?"
- "Which signals are critical today, and what does each one say?"
How does Reperline handle Claude access?
Reperline treats Claude as one more user with limited rights. Connections and changes are kept in an activity log that nobody can edit or delete from the product, and passwords and tokens are never recorded. Claude never writes to the database directly, and it cannot approve its own proposals, because approval is a separate action taken by a person in the Inbox.
When a question asks for something the product does not calculate, such as earned value or CPI, Claude should say so rather than invent a number. An answer with no figure behind it is better than a made up one.
The owner can see who connected and who changed what in the activity log, so a question about where a figure came from has a recorded answer. Rights are resolved on every call, so a role that is changed or removed applies to the next request, not only to the next sign in.
What Claude can do is also bounded by the product. It cannot produce a forecast that the product does not calculate, such as an earned value EAC, and it cannot read a file the connector does not serve. If a figure is missing from the system, the honest answer is that it is missing. The audit module describes the log, and the budget module explains which figures Claude can propose.
Claude in the office is included on every plan, with unlimited users, so a plan choice does not change who can connect. The prices are on the pricing page. For the signals Claude can list, read the early warning article.