Deploy from Claude Code or Codex: MCP setup
Connect Agent Lab's MCP server to Claude Code, Codex, Claude Desktop or another MCP client and let it create the project, upload and go live. Sign in once.
1. Connect Agent Lab MCP#
Add the connection in your tool: Claude Code and Codex take a terminal command, Claude Desktop a settings screen. Replace acme with your organization's name; the console's MCP page shows the command, and the connection name agentlab-<organization>, with it filled in.
claude mcp add --transport http --scope user agentlab-acme https://mcp.agentlab.cresclab.com/mcpAdd a custom connector in Claude Desktop (or claude.ai on the web). On a Team or Enterprise plan, an organization owner adds it first in Organization settings › Connectors (Add › Custom › Web); members then choose Connect in Customize › Connectors, sign in and pick their organization in the browser, and continue from step 4. The Free plan allows one custom connector.
- Open Customize › Connectors (Settings › Connectors in older versions) and choose + Add › Add custom connector.
- Name it
agentlab-acme, enterhttps://mcp.agentlab.cresclab.com/mcpas the URL and choose Continue, then Continue again on the screen that shows the detected authentication settings. - Under Authentication choose Sign in now, keep Use Claude's published identity (Recommended) as the OAuth client (not Register automatically, which Agent Lab does not support yet), choose Add, then sign in and pick your organization in the browser.
- Start a new conversation, open + › Connectors at the lower left of the message box and turn on
agentlab-acme. - In that conversation, ask "List my Agent Lab projects". If Claude calls
sites_listand lists your projects, you are connected. If the tools do not appear, fully quit and reopen Claude Desktop and try again.
A regular Claude Desktop chat cannot read folders on your computer. It can check status, change settings and put a version live. To upload a local project, use Claude Code (a terminal, or the Code tab of Claude Desktop); to deploy from GitHub, connect the repository in the console.
codex mcp add agentlab-acme --url https://mcp.agentlab.cresclab.com/mcp
codex mcp login agentlab-acmeOr add it to ~/.codex/config.toml, then run codex mcp login agentlab-acme:
[mcp_servers.agentlab-acme]
url = "https://mcp.agentlab.cresclab.com/mcp"Add an HTTP MCP server at:
https://mcp.agentlab.cresclab.com/mcpThe client needs OAuth 2.1 with PKCE, and identifies itself with a Client ID Metadata Document or by dynamic client registration, so Cursor, VS Code, Windsurf, Gemini CLI, Qwen Code and similar clients work. Sign-in opens in the browser. A client that does no OAuth and can only send a fixed header cannot connect yet.
A sign-in page opens in your browser. Sign in and choose your organization to finish.
No password, token or API key
Authorization happens in the browser. The connection can do what you can do in the console, and nothing more.
When you add the connection before opening Claude Code, as above, the tools work as soon as you finish signing in from /mcp; there's nothing to restart. Only a connection added while a Claude Code session was already open is missing from that session's /mcp: exit and run claude --continue in the same folder, which keeps the conversation. A Claude Desktop connector works in a new conversation with it turned on; only if its tools don't appear, fully quit and reopen Claude Desktop. In Codex, sign in before you start the session, or start a new one after. If the tools look out of date later, reconnect the server (in Claude Code, /mcp › the server › Reconnect); that doesn't need a new sign-in.
The sign-in allows projects, databases and caches together, and each only as far as your own role in the organization allows. A tool that asks for less gets less.
Organization owners and admins, and a database's managers, already have writer access without a separate grant; only other members need a reader or writer grant. Once you approve the SQL, the assistant should call the tool directly and ask for an access change only after an actual denial. For a project that needs tables, it creates them with databases_apply_migration, then sets up the project's credentials with databases_connect_app. You don't need to create tables by hand in the console.
2. Ask it to deploy#
In your project folder, ask the tool to read the deploy guide and follow it:
Read https://app.agentlab.cresclab.com/agent/deploy.md and follow it to deploy this folder to the Agent Lab organization acme.If the tool can't read that address, the same guide is at https://agentlab.cresclab.com/agent/deploy.md. Once Agent Lab MCP is connected, docs_read also returns these docs.
To have it look something up in these docs, give it a page's Markdown address or the index at https://agentlab.cresclab.com/docs/en/llms.txt. The MCP server also has a docs_read tool that returns the same pages.
To make this the default, put the same sentence in the project's AGENTS.md (Codex) or CLAUDE.md (Claude Code):
When asked to deploy, read https://app.agentlab.cresclab.com/agent/deploy.md and follow it to deploy to the Agent Lab organization acme. If that address can't be read, read https://agentlab.cresclab.com/agent/deploy.md or call docs_read.Using ChatGPT or Claude on the web?
Those tools can't read files on your computer. Save the project as a zip and upload it on the project's page in the console, or deploy from GitHub.
3. What it does#
- Finds or creates the project It confirms the address label with you first, because labels can't be reused.
- Uploads It uploads the zip straight to the address the platform gives it. Your files don't pass through the conversation.
- Waits for the build It checks the version's status until the build succeeds or fails.
- Asks before going live It tells you which version will go live and waits for your OK.
You still enter secrets yourself
The tool tells you where to enter a secret. It won't ask you to paste one into the conversation.
See MCP tools for every tool. When the code needs PostgreSQL, MySQL or Redis, the tool creates and connects it too: see Databases and Redis caches.
4. What connecting changes on your computer#
Connecting adds one server entry to your tool's MCP settings and saves one sign-in. That is all it writes.
| Tool | What is added | Where |
|---|---|---|
| Claude Code | An agentlab-acme entry under mcpServers | ~/.claude.json, for you in every folder (--scope user, as in step 1). Without --scope, the same file, for the current folder only. --scope project would write .mcp.json into the project and share it through git; we don't suggest it. |
| Claude Code | The sign-in | The macOS Keychain; on Linux and Windows, ~/.claude/.credentials.json |
| Codex | An [mcp_servers.agentlab-acme] table | ~/.codex/config.toml |
| Codex | The sign-in | Where Codex keeps MCP sign-ins (the system keyring or a file, by its mcp_oauth_credentials_store setting) |
| Claude Desktop, claude.ai | A custom connector | Your Claude account, not a file on your computer. claude_desktop_config.json is not changed. |
| Other clients | One server entry | The client's MCP settings |
What it doesn't change
- Your skills, slash commands, hooks, plugins, subagents and other MCP servers. Agent Lab installs none of them.
settings.json, permissions and the model you use. If you answer a tool prompt with "don't ask again", your tool records that, the same as for any other tool.CLAUDE.mdandAGENTS.md. The line in step 2 is optional, and you add it yourself if you want it.- Nothing is installed or runs on your computer: the server is remote, so there is no package,
npxcommand or background process.
Its tools carry the connection's name, such as mcp__agentlab-acme__sites_list in Claude Code, so they can't take the place of a skill or another server's tool with the same name. Your agent uses them when you ask it for something in Agent Lab.
To undo it: in Claude Code, claude mcp remove --scope user agentlab-acme, and in /mcp, Clear authentication first if you also want the saved sign-in gone. In Codex, codex mcp remove agentlab-acme. In Claude Desktop, remove the connector under Customize › Connectors. To end the connection on Agent Lab's side as well, remove it on the console's Deploy with an AI tool page, under connected applications.