The MCP door

The MCP door lets your organisation’s own AI agent read the ledger. The agent speaks the Model Context Protocol (MCP), revision 2026-07-28, which most agent tools already speak. It sees the same lists a person with the same permissions sees, and nothing more.

A controller at Woodgrove Bank can ask their agent to reconcile a GL account against the ERP. The agent lists what it may read, reads the journal entries, and cites the ids it read. Every read is kept, and a checker can review it later.

What the door does not do

Who can use the door

Two kinds of caller use the door.

Caller How they sign in What they may read
A person Through your organisation’s usual sign-in, in their agent tool What their own permissions allow
An agent With its own client id and secret, with no person present What its group’s permissions allow

An agent is a member of your organisation, in its own group. Give it the group a clerk in the same job would have. An agent can never approve anything, even if its group allows approvals.

Each caller also needs the permission to use the MCP door.

Connect an agent tool as a person

  1. In your agent tool, add a remote MCP server.
  2. Enter the door’s address: your platform address, then /mcp/, then your organisation’s key. For example, https://ledger.woodgrove.example/mcp/woodgrove.
  3. Connect. The tool opens your organisation’s sign-in page.
  4. Sign in as you usually do.

The agent tool finds the sign-in page by itself. If you are not a member of the organisation in the address, the door refuses you.

If the MCP door is switched off, the door refuses every request. The refusal names MCP Door and says an administrator can switch it on in Settings → Features. If your licence does not include the MCP door, the refusal says so.

Connect an agent that runs on its own

An agent that runs with no person present needs its own credential.

  1. In your identity provider, register a client for the agent with the client-credentials grant. Give it access to the audience ledgertm-mcp.
  2. Note the client id and create a secret. Keep the secret in your own vault.
  3. Create a user group for the agent in Permissions, and grant it what it may read.
  4. Ask your platform administrator to register the client id as an agent for your organisation and that group.
  5. In the agent’s configuration, enter the door’s address, the client id and the secret.

The platform administrator’s registration creates the agent’s member, named by its client id. A client id already used by an institution connection is refused. Register a new one.

Revoking the agent is also a platform administrator’s task. It takes effect within about a minute.

What the agent can see

The agent asks the door for its tools and its resources.

What What it is
A tool for each list Every list its permissions allow, such as journal entries or account balances
find_tools Searches those lists by words, such as journal
aggregate Groups a list’s rows by a field and totals them, as the assistant does
run_report_sql Runs one SQL SELECT over your reporting tables. Shown only when your licence includes Reporting tables and the agent holds its permission
Guide articles Every public article in this guide, as text
The reporting dictionary Every reporting table and column, with its description. Shown only when your licence includes Reporting tables

SQL reads the live book only. There is no “as at” date; for an earlier date, read the day-end tables that carry one. Each statement stops after a time limit and returns a capped number of rows.

Review what an agent read

Every tool call and every resource read is kept. One caller’s calls on one trading date are one conversation.

  1. Go to the Assistant conversations screen.
  2. Set Source to MCP door.
  3. Select a conversation to open it.

You see each call, its parameters, how many rows came back, why it was refused if it was, and the trading date. See The assistant for the columns.

What your organisation configures

From the LedgerTM product guide, as shipped with the software and written with it. Copied from the platform repository on 2026-11-11.