AI agents (MCP)

Let an AI coding agent such as Claude Code or Codex draft content and propose contract changes, while people keep control of publishing.

Atlas exposes a Model Context Protocol endpoint so an AI coding agent working in your website’s repository can do the repetitive parts of a migration: model the site contract, upload images, create entries and pages, and draft translations.

What an agent can and cannot do

Agents can	Agents can never
Read the contract; validate and propose changes	Approve or activate a contract
Create and edit drafts of pages, entries and globals	Publish, in any environment
Draft translations	Roll back or delete
Upload media	Touch Production in any way
Create and edit chatbot drafts	Run shell commands, SQL, file access or fetch URLs
	Deploy anything

A person reviews every draft in Studio and publishes it.

Connect an agent

In Studio open Integrations → Coding agents → New connection. Choose the Development environment and the permissions (grants) the agent needs:
- contract:read, contract:propose
- content:draft
- media:upload
- bots:draft
- plans:read (Read website plans: approved helper blueprints, palettes and design tokens; new in this release)
Only site and organization admins can create connections. Click Create connection.
Copy the token. It starts with amc_ and is shown once.

Add the server to your agent. The endpoint uses Streamable HTTP:

{
  "mcpServers": {
    "atlas": {
      "type": "http",
      "url": "https://mcp.atlas.example/mcp",
      "headers": { "Authorization": "Bearer ${ATLAS_MCP_TOKEN}" }
    }
  }
}

Keep the token in an environment variable, never in a committed file.

Ask the agent to call atlas.connection.get. It should report your site and its Development environment.

Studio’s Guides contain ready-made briefs for each migration step, with your site’s values filled in.

Safety controls

Connections can only be created for Development environments and expire within 90 days.
Connections made before plans:read existed must be recreated to get it.
Each call is checked against the live connection, so revoking a connection takes effect on the next call.
Calls are rate limited (120 a minute by default), and every write needs an idempotency key.
Agent actions are recorded in the audit log.

Tools

Grant	Tools
(always)	`atlas.connection.get`
`contract:read`	`atlas.contract.get`, `atlas.contract.list`, `atlas.contract.proposal.get`
`contract:propose`	`atlas.contract.validate`, `atlas.contract.propose`, `atlas.contract.proposal.withdraw`
`content:draft`	`atlas.pages.list`, `.get`, `.createFromType`, `.sections.update`; `atlas.entries.list`, `.get`, `.upsertDraft`; `atlas.globals.list`, `.get`, `.updateDraft`; `atlas.translations.get`, `.updateDraft`
`media:upload`	`atlas.media.list`, `atlas.media.asset.get`, `atlas.media.uploadSlot.create`
`bots:draft`	`atlas.bots.list`, `.get`, `.create`, `.updateDraft`
`plans:read`	`atlas.helper.projects.list`, `atlas.helper.blueprint.get`, `atlas.brand.palettes.get`, `atlas.brand.designTokens.get` (read-only)

There are no agent tools for fonts, the library, animations or design scans, and no tool can publish or approve anything.