Esc

↑↓ move↵ openIndex · Pagefind
Technology

Model Context Protocol (MCP)

Definition

An open protocol that lets an AI application discover and call tools and read data from other systems through a standard connector.

Topic: Technology & data

Updated

The Model Context Protocol (MCP) is an open standard for connecting AI applications to the systems where work happens. An MCP server offers a set of operations, such as “list open RFIs” or “read a commitment,” with a description and an input schema for each. An MCP client, typically an AI assistant, discovers those operations and calls the ones it needs to answer a question or carry out a task.

For construction, the point is to let an assistant work from the job’s actual records instead of pasted screenshots and stale exports, without giving it a back door around the permissions those records already have.

How it works

  1. A server describes its operations. Each has a name, a description and the inputs it accepts.
  2. A client connects with a credential that identifies a person and what they’re allowed to do.
  3. The assistant discovers operations relevant to the request and reads the schema for the one it picks.
  4. It calls the operation. The system behind the server checks the credential and the person’s permissions on every call.
  5. Results come back bounded, a page at a time, for the assistant to use or summarize.
  6. Calls are recorded, so there’s a trail of what was read or changed and by whose credential.
What it is Who uses it
API The system’s programmatic interface Developers and integrations
MCP server A connector that presents operations to AI clients AI assistants and agent tools
Webhook A notification the system sends when something changes Integrations that react to events
AI agent Software that plans and carries out steps toward a goal Teams delegating repetitive work

Why it matters

Most of a contractor’s questions cross records: which RFIs are holding up change orders, which commitments are over budget, which subs have expiring insurance. An assistant connected through MCP can answer from the source, but it inherits whatever the credential can do. That makes the governance questions the same ones you’d ask about a new employee’s login: what can it see, what can it change, and who approves.

Worked example

Illustrative: a project manager connects an assistant with a token scoped to read RFIs and change events on one project, writes disabled. She asks for every open RFI older than ten days that’s linked to a change event. The assistant searches the available operations, reads the RFI list a page at a time, cross-checks the linked change events, and returns six RFIs with their ball in court. When she asks it to close one, the call is refused, because the connection is read-only. She closes it herself in the app.

Common mistakes

  • Connecting with a shared administrator credential instead of the person’s own.
  • Granting write access “just in case.”
  • Pasting API tokens into a chat prompt or a shared configuration file.
  • Treating the list of operations a server describes as a list of what the user is allowed to do.
  • Letting an assistant approve financial records instead of drafting them for a person.

How os.construction handles it

An MCP connector and an agent CLI are in Preview as a local developer preview: they run from a checkout of the os.construction repository and call the REST API with the person’s own API token, so the API’s normal checks on scopes, company, project membership and enabled tools apply to every call. The connector keeps its tool list small: search operations, describe one, then call it. Writes stay off unless explicitly enabled. Hosted connections with browser consent, listed and revocable under the account’s connected apps, have passed local checks but aren’t publicly released. Built-in AI agents remain planned. See Set up the MCP preview and MCP permissions and troubleshooting.

FAQ

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open protocol for connecting AI applications to external tools and data. An MCP server describes the operations it offers, and an MCP client, such as an AI assistant, can discover those operations, read their input schemas and call them. The specification is published at modelcontextprotocol.io.

Is MCP the same as an API?

No. An API is the interface a system exposes, usually over HTTP. An MCP server is a connector that presents operations, often backed by an API, in a form AI clients understand. The API's own authentication and permission checks should still decide what each call can do.

Is it safe to connect an AI assistant to project and financial data?

It can be, if the connection uses the person's own credential with the narrowest scopes the task needs, the system enforces its normal permissions on every call, writes are off unless deliberately enabled, and every call is recorded. Treat a connection like giving someone a login: decide what it may see and change before you connect it.

Can an AI assistant approve a change order or payment through MCP?

Only if the credential and the system allow that user to do it. Good practice is to keep approvals with people: let the assistant read, summarize and draft, and require a named person to approve anything that commits money or changes a contract.

Sources

Open file