MCP

The Major MCP server exposes Major itself as Model Context Protocol tools. Connect an MCP client and your AI assistant can list and create apps, agents, skills, and workflows, start sandboxes and edit their files, query your connectors, run agents and workflows, and deploy - from the client you already use. It gives your client the same capabilities as the Platform Agent in the web app.

Example requests

“List the apps I can edit in Major.”
“Create an app that shows open support tickets from our Postgres database, then deploy it.”
“Add a status filter to the dashboard in my ops-dashboard app.”
“Build an agent that summarizes yesterday’s Slack messages in #sales.”
“Run the weekly-report workflow with the input { \"region\": \"EMEA\" }.”
“How many customers signed up last week?”

Connect

The recommended way to connect is the Major plugin. It configures the MCP server and installs the Major skills in one step. See the Quickstart for install steps for each client.

To configure the server by hand, use:


Endpoint	`https://mcp.major.build/mcp`
Transport	Streamable HTTP
Authentication	OAuth

The first time your client connects, a browser tab opens to Major. Sign in and click Allow to grant the client access to your organization. The client then stores the token and connects without prompting again.

The access token is tied to one organization. To work in a different organization, reconnect the server and approve access for that organization.

Skills in the plugin

The plugin also installs skills that teach your assistant how to build on Major. The assistant loads them on demand:

Skill	Use it for
`new-project`	Planning a brand new app: connector setup and the data layer
`app-builder`	Creating and editing apps in an app sandbox
`agent-builder`	Creating and configuring agents (`agent.jsonc` and `prompt.md`)
`skill-builder`	Creating and managing agent skills
`workflow-builder`	Creating and managing workflows, their nodes, and triggers
`using-connectors`	Querying connectors and Major platform resources, with per-connector references
`using-apps`	Calling the API of a deployed app
`using-authn`	Adding sign-in and identity to an app
`using-webhooks`	Receiving webhooks in a deployed app
`using-frontend-data-table`	Building table views in an app’s frontend
`debug-issue`	Investigating errors, failed calls, and broken behavior

The Platform Agent uses the same skills, so your client and the web app build things the same way.

What it can do

Group	Tools	What they do
Discovery	`list`, `list_sandboxes`	List the apps, agents, skills, or workflows you can edit (or view, with `include_read_only`). List your running sandboxes.
Create	`create`	Create an app, agent, skill, or workflow and start its sandbox.
Sandboxes	`start_sandbox`, `stop_sandbox`	Start the sandbox for an existing app, agent, skill, or workflow, or release it.
Sandbox editing	`sandbox_read_file`, `sandbox_write_file`, `sandbox_edit_file`, `sandbox_glob`, `sandbox_grep`, `sandbox_bash`, `sandbox_bash_output`, `sandbox_bash_stop`, `sandbox_workspace_status`, `sandbox_restart_preview`	Read, write, and search files, and run shell commands in a sandbox. `sandbox_restart_preview` restarts an app’s preview server.
App wiring	`sandbox_add-resource-client`, `sandbox_remove-resource-client`, `sandbox_add-agent-client`, `sandbox_remove-agent-client`	Add or remove a connector or an agent client in an app.
Connectors	`search_resource_tool_catalog`, `execute_resource_tool`	Find a connector tool by describing what you need, then run it against one of your connectors.
Connector setup	`search_connector_types`, `request_resource_setup`	Find connector types and get a prefilled link to set up a new connector.
Deployed apps	`get_app_skill`, `do_get_request`, `do_requests`	Read a deployed app’s API guide and call its endpoints. Redirects from the app aren’t followed: the call returns the `3xx` status and its `Location` header.
Agent logins	`request_agent_login`	Get a link where someone with edit access to an agent saves a website login for it, so the agent’s browser can sign in to that site.
Running	`run_agent`, `run_workflow`	Start a run of an agent’s published version, or run the latest saved draft of a workflow.
Environments	`get_environment`, `set_environment`	See or switch the environment your connector calls and app previews use. Only shown in organizations with more than one environment.
Data and config	`setup_managed_database`, `run_migration`, `set_app_env_variables`	Create an app’s managed database, run SQL migrations on it, and get a link to set an app’s environment variables.
Access	`list_subjects`, `list_access`, `set_access`, `list_subject_access`	Find users and groups, see who has access to an app, agent, skill, workflow, or connector, and grant, change, or remove a role on it. See Managing access.

Tools that need secrets or credentials, such as request_resource_setup and set_app_env_variables, never take them from your assistant. They return a link that you open in Major to enter the values yourself.

How it works

Sandboxes

Apps, agents, skills, and workflows are all edited on a remote sandbox. Your assistant finds what to edit with list, starts its sandbox with start_sandbox (or create for something new), and then edits through the sandbox_* tools. Every sandbox call names its target: an app’s slug, or an agent, skill, or workflow id. Several sandboxes can run at once. You can watch every sandbox your client starts at app.major.build/sandboxes, with one tab per sandbox, including an app’s live preview. A sandbox is held by one person at a time. If someone else is editing the same item, start_sandbox reports that it is locked.

Saving, publishing, and deploying

Inside a sandbox, sandbox_bash runs the Major CLI in the workspace root:

Command	Agent, skill, or workflow	App
`major pull`	Pull the saved bundle	`git pull`
`major push`	Save a new version	`git push`
`major validate`	Validate the bundle	Run lint
`major publish`	Publish the latest saved version	Deploy the pushed commit

An app’s first deploy needs major publish --slug <slug> to choose its URL. See the CLI for the full command set.

Connectors

Your assistant does not load a separate tool for every connector. It describes what it needs to search_resource_tool_catalog (for example, “read-only SQL against Postgres”), gets back matching tools with their input schemas, and runs one with execute_resource_tool. See Connectors.

Managing access

Ask your assistant who can reach something, or to share it, for example “Who has access to the ops dashboard?” or “Make the Support group an Editor on the triage agent.”

Tool	Who can use it	What it does
`list_subjects`	Builders and admins	Find users or groups by name or email, or list a group’s members.
`list_access`	Builders and admins	List the users and groups that hold a role on one app, agent, skill, workflow, or connector.
`set_access`	Builders and admins	Grant, change, or remove a user’s or group’s role on one object.
`list_subject_access`	Organization admins	List everything one user or group can reach, directly and through groups.

list_access and set_access only work on objects whose access you can manage, meaning you’re their Admin or an organization admin. They follow the same rules as the share dialogs in the web app. A role must match the object type. Members and external users can only get User roles. An object always keeps at least one Admin unless an organization admin removes it. Removing a user’s direct role doesn’t remove access they have through a group. The Platform Agent has the same tools. There, set_access asks for your approval before it runs.

What `execute_resource_tool` and `sandbox_bash` can reach

These two tools are broad, so here is exactly what each one can touch.

`execute_resource_tool`

execute_resource_tool runs one tool from your organization’s connector catalog. It reaches only connectors that an admin has configured in Major, and of those, only the ones your roles let you use. It cannot reach any system that isn’t set up as a connector.

Connector type	Examples	Write actions it can take
Databases	Postgres, MySQL, SQL Server, Snowflake, BigQuery, ClickHouse, DynamoDB	SQL that inserts, updates, or deletes rows or changes the schema
Internal APIs	Your organization’s REST and GraphQL APIs	`POST`, `PUT`, `PATCH`, and `DELETE` requests, and GraphQL mutations
SaaS connectors	HubSpot, Salesforce, Gmail, Slack, Google Drive, Notion, GitHub	Creating, updating, sending, or deleting records, such as sending an email or a Slack message
Storage	S3 and managed file storage	Uploading, overwriting, and deleting files

Whether a write succeeds also depends on the connector’s own credentials. A database connector set up with a read-only user can only read. See the connector reference for every connector type. For PostgreSQL, MySQL, SQL Server, and ClickHouse, a query’s result returned to the AI is capped at 1 MB. A larger result fails with result exceeds 1 MB; add a LIMIT or select fewer columns. Add a LIMIT, select fewer columns, or page through the rows. Queries that apps run through the resource client aren’t affected by this cap.

`sandbox_bash`

sandbox_bash runs shell commands inside the isolated edit sandbox of one app, agent, skill, or workflow. The sandbox is a dedicated container that holds only that item’s workspace.

Network scope: public internet only. Major’s private network and the cloud metadata endpoint are blocked.
Connectors: reachable only through Major’s API, using the same access your roles give you. The sandbox holds no connector credentials.

Running major publish from sandbox_bash ships to production. For an app, it deploys the pushed commit to the app’s live URL. For an agent, skill, or workflow, it makes the latest saved version the live one. Your client runs it like any other shell command, so use your client’s tool approval settings if you want to confirm each sandbox_bash call.

Permissions

The MCP server acts as you:

Scoped to your organization. Every call runs in the organization you approved when you connected.
Scoped to your roles. list returns only what you can access, and each tool checks your access to the item it touches. Creating things, starting sandboxes, running agents and workflows, and connector setup require permission to build in the organization. Members without it see only the connector, deployed-app, and listing tools. See Members & roles.
Approvals happen in your client. Your personal tool permissions apply to Platform Agent chats in the web app, not to MCP clients. Use your client’s own tool approval settings to control what runs without asking.

Next steps

Quickstart

Install the Major plugin in your client.

Platform Agent

The same capabilities, built into the web app.

CLI

Build on Major from your own terminal.

Get Started

Learn

Build

Settings

Reference

Example requests

Connect

Skills in the plugin

What it can do

How it works

Sandboxes

Saving, publishing, and deploying

Connectors

Managing access

What `execute_resource_tool` and `sandbox_bash` can reach

`execute_resource_tool`

`sandbox_bash`

Permissions

Next steps

Quickstart

Platform Agent

CLI

Example requests

Connect

Skills in the plugin

What it can do

How it works

Sandboxes

Saving, publishing, and deploying

Connectors

Managing access

What execute_resource_tool and sandbox_bash can reach

execute_resource_tool

sandbox_bash

Permissions

Next steps

Quickstart

Platform Agent

CLI

What `execute_resource_tool` and `sandbox_bash` can reach

`execute_resource_tool`

`sandbox_bash`