Junction

Reference

Junction MCP tools: the six tools every AI assistant sees

Updated 11 November 2026

Short answer: Junction’s MCP server offers exactly six tools to Claude, ChatGPT, Gemini and Copilot: list_connected_systems, describe_entity, query_entity, fetch_more, get_skill and save_skill. Five only read. The sixth saves a skill a person asked for in Junction’s own skills library and never writes to your systems. Systems and what can be asked about each are data the tools return, so connecting a new system adds no tool.

What are the six tools?

Each tool has a title. Five are flagged to the AI app as ones that change nothing; save_skill is flagged as the one that is not read-only, because it saves a skill, and it changes nothing in any system you connect. The table lists them in the order an assistant usually uses them.

ToolWhat it doesInput
list_connected_systems
List connected systems
Returns the systems the signed-in company has connected, with the entities each one can be asked about, and the saved skills the person may use. Another company's systems never appear.None.
describe_entity
Describe what can be queried
Returns, for one system, each entity's fields (name, type, meaning) and measures. Anything the person's role does not allow is left out.The system.
query_entity
Query connected systems
Asks one system for rows or totals as a structured request: fields, filters, groupings and measures. The assistant picks from named fields and the values travel as bound parameters, so a question is never run as a command.The system, the entity, and the fields, filters and measures wanted.
fetch_more
Fetch the next page of results
Returns the next page of a large answer. The cursor works only for the person and company that received it, and the person's limits are applied again when it is used.The cursor returned with the previous page.
get_skill
Read a saved skill
Returns one saved set of instructions the person is allowed to use, such as how a company wants a job cost review written. A skill adds no data access.A skill id from list_connected_systems.
save_skill
Save a skill
Saves a set of instructions the person asked for, for themselves or for their company, so it is offered again in later conversations. It can only create skills and change the ones that person created, and every change is kept as a version. It never writes to a connected system and adds no data access.The skill's name, when to use it, its instructions and who it is for; to change one, its id.

How does a question flow through the tools?

The assistant starts with list_connected_systems, reads what a system offers with describe_entity, asks with query_entity, and calls fetch_more if the answer has more pages. When a saved skill fits the question, it reads it with get_skill, and when a person asks it to keep a skill it calls save_skill. Each answer carries its source and the time the data was current.

How do permissions apply to every tool?

Every call is checked against the signed-in person’s company and role. An administrator can limit a person by system, table, column, row or sensitivity label. Access can only be narrowed, never widened, and anything a person may not see is also left out of describe_entity. The audit log records who asked and whether it worked, never the values.

What is a skill?

A skill is a saved set of instructions, such as how your company wants a job cost review written. People keep one private, or an administrator shares it with the company or with chosen roles. Reading one through get_skill adds no data access, and it is served with a note that it cannot override the user. A person can ask their AI app to write a skill and save it with save_skill, for themselves or for their company; sharing one with other companies is done in the portal and reviewed first. Anyone who can change a skill can switch it off, each person can hide a skill from themselves, and every change is kept as a version that can be restored.

Frequently asked questions

What tools does the Junction MCP server offer?
Six: list_connected_systems, describe_entity, query_entity, fetch_more, get_skill and save_skill. The list never grows with each system, because systems and their entities are data the tools return, not new tools.
Can any Junction tool change data?
Not in your systems. Five of the tools only read and are flagged to the AI app as ones that change nothing. The sixth, save_skill, saves the text of a skill that a person asked their AI app to keep, in Junction's own skills library, and is flagged as one that is not read-only. Junction never writes, edits or deletes anything in the systems you connect.
Why does Junction have so few tools?
AI apps load every tool definition into the model's context, and some freeze the list when a connector is published. A short, fixed list keeps answers faster and means a new system needs no change in your AI app.
How does Junction decide what a person can see through these tools?
Every call checks the person's company and role. An administrator can limit a person by system, table, column, row or sensitivity label, and access can only be narrowed, never widened.