MCP tools

This page is generated from the code of engine version 0.16.0. Between releases, the site shows the code on main, which may already have changes for the next version.

Pointsman serves the Model Context Protocol (MCP) at /mcp, so AI agents can ask a decision profile instead of deciding on their own. MCP server explains how to connect.

list_profiles

List the decision profiles you can use: their questions, answer options and possible actions.

No arguments.

explain_profile

Explain a decision profile in plain sentences: what the AI is asked, which facts and earlier decisions it uses, and each rule as “If …, the action is …”. Show the text to the person who approves the profile.

Argument	Type	Required	Description
profile	string	yes	Profile id from list_profiles
lang	“en” | “ja”	no	Language of the sentences (default en)
version	integer	no	A published version (default: the latest)

replay_profile

Try a changed policy on the decisions already made: what the new rules would have decided on the last logged decisions of a profile version, with the logged answers and facts (no model is called). Returns how many actions would change, from what to what, how often they would match what a person answered, and the changed decisions. Only the policy may change; send questions, facts and requires too to check that they are the same.

Argument	Type	Required	Description
profile	string	yes	Profile id from list_profiles
policy	object	yes	The new policy: { rules: [{ when, action }], default, people_check?, people_answers? }
version	integer	no	The published version to replay (default: the latest)
limit	integer	no	How many of the newest decisions (default 200)
questions	array	no	The draft’s questions, to check they did not change
facts	array	no	The draft’s facts, to check they did not change
requires	array	no	The draft’s requires, to check they did not change

decide

Ask a decision profile. Returns typed answers with probabilities and the action to take (for example “auto”, or “review” when a person should decide). Use list_profiles first. For profiles with input fields, send state as an object with those fields under the names the profile expects.

Argument	Type	Required	Description
profile	string	yes	Profile id from list_profiles
state	string | object	yes	What to decide about: text, or a JSON object
ref	string	no	Your reference for this decision, for example an issue URL
decisions	object	no	For profiles with requires: the decision_id of each earlier decision, by the name in requires. Needs ref.
idempotency_key	string	no	Send the same key again to get the same decision instead of a new one

get_decision

Get a decision by id: answers, action, profile version, model, review status and feedback.

Argument	Type	Required	Description
decision_id	string	yes	decision_id returned by decide

submit_feedback

Correct a decision: give the right answer for one or more of its questions. Recorded in your name and used to measure how often the profile is right.

Only for people logged in through OAuth, not for API tokens.

Argument	Type	Required	Description
decision_id	string	yes	decision_id returned by decide
correct	object	yes	The right answers, by question name, for example {“team”: “frontend”}
note	string	no	Why, in a sentence