Margince: a CRM where AI agents work under the same rules as users (MCP/REST access)

Who it's for

Margince is a CRM that fills in its own data: the official page states that it auto-fills its own records from mail, meetings and calendars, and that AI agents work in it under the same rules as its users. It suits two kinds of people: teams and individuals whose customer information is scattered across inboxes, calendar invites and meeting threads, and who want to keep manual entry to a minimum; and those who want AI agents to take on everyday actions such as lookups and follow-ups directly in the CRM, without the agents bypassing the human permission system.

Cases that do not fit, or that deserve a second thought first: if you do not want to run a service yourself — the official page only says the core is a Go module in backend/ and a React app in frontend/, which you build and run from source yourself, and it names no specific dependency tools or versions; and if you expect a ready-to-use product before obtaining a license token — the official page states explicitly that the server needs a license token from Gradion to start, even for a single seat.

Your first attempt in 3 minutes

The only main path the official page supports is this: get the margince source from the official repository, enter the repository root, and run:

make install

The official wording is “Install the tools and the git hooks once: make install.” — a one-time installation of the required tools and git hooks. Observable success signals: the tools are installed and the repository's git hooks are active. This is the full extent of the installation steps the official page supports.

You must know this before you start: starting the server requires a license token issued by Gradion, even for a single seat; the official page mentions no development exemption or free tier, so do not assume the local service can run without a token. Until you have a token, installation is as far as you can get.

Beyond that, the official page says access is over MCP or REST, but the startup command, the service endpoints and the names and parameters of the agent tools are not given there, so verify them yourself in the repository. Once connected, keep your first call to read-only or discovery actions, see what the agents can do, and only then consider anything that modifies records.

Use cases

Case 1: let an agent browse customer records within permission rules. Scenario: verify that AI agents can really work in the CRM. How: after completing the installation and sorting out the license token, connect an MCP client to the service and send one lookup request. Expected result: the official positioning is that agents follow the same rules as users, and what an agent can see is decided by its permissions. Evidence boundary: the official page gives no specific tool names or endpoints; what can be confirmed is that access is over MCP or REST — check the tool list in the repository before sending requests.

Case 2: let the CRM auto-fill customer records from mail, meetings and calendars. Scenario: reduce manual entry by turning customer interactions scattered across the inbox and calendar into records. How: once the service is available, connect these data sources. Expected result: records are filled in automatically by the system, the core behavior described on the official page. Side-effect warning: auto-filling writes information from mail, meetings and calendars into CRM records; the page does not explain how data sources are authorized to connect, so decide before connecting which mailboxes and calendars to link and which not.

Usage boundaries

Read/write nature: Margince automatically writes information from mail, meetings and calendars into CRM records, and once the data is in, it becomes record content; agents follow the same rules as users with comparable capability ceilings, so it is safer to keep the first attempt to reads and discovery.

Licensing and credentials: the server needs a Gradion license token to start, even for a single seat; the official page says nothing about a development exemption, a free tier or pricing, so do not assume local development is license-free.

Key unconfirmed details: dependency tools and versions, the startup command, MCP/REST endpoints, the agent tool list, the license application process and whether a hosted version exists are not explained on the official page. The official information is very limited, and this article only covers what it supports; for any precise configuration the actual files in the repository are authoritative — check each item before you start.