Valet|Docs

MCP Server

Valet runs its own MCP server. Point any MCP client at it and that tool can publish artifacts, list them, change who can see them, and take them down — without a terminal, an install, or an account.

https://api.valet.dev/mcp

This is the path for the tools where the Valet CLI cannot run: Claude Cowork sandboxes outbound traffic, and claude.ai, ChatGPT, and mobile have no shell at all. A remote MCP connector is dialed from the vendor's cloud rather than from the sandbox, so it reaches Valet where the CLI does not.

Use the CLI when you can

If your tool can run commands — Claude Code, Codex, Cursor — install the Valet skill instead. It is the fuller surface: it publishes whole directories from disk, downloads an artifact to edit and redeploy it from anywhere, sets password access, and manages agents, integrations, channels, and members. The MCP server publishes the files a tool hands it and does the five things below. Reach for it when the CLI is not an option, not before.

Add it to your tool

Connect Your Agent covers every client in one place, including the ones that install the skills instead. The connector steps are repeated here for reference.

Claude Cowork and claude.ai

Settings → Connectors → Add custom connector. Paste the URL and connect. Claude registers itself, so there is no client ID or secret to create.

ChatGPT

Custom connectors sit behind developer mode. Settings → Connectors → Advanced → turn on Developer mode. Then add a custom connector and paste the same URL. Developer mode needs a paid plan — Plus, Pro, Business, Enterprise, or Edu. Free accounts cannot add custom connectors.

Claude Code

Available, though the skill is the better fit here — see above.

claude mcp add --transport http valet https://api.valet.dev/mcp

Anything else

The server speaks Streamable HTTP and advertises its own authorization metadata, so a client that implements MCP needs nothing but the URL.

Signing in is optional

The server has three modes, and a client picks one by what it presents.

  • Nothing at all. Publishing works with no account. The artifact is public to anyone with the link and is removed 36 hours after it is created. The result carries a claim URL that makes it permanent, and a handle the tool reuses to update or delete that one artifact later in the conversation.
  • Signed in. Connect the account when your client offers it and artifacts publish into your organization: permanent, private on creation, and visible in the dashboard beside everything else your team has shared. Authorization uses OAuth with dynamic client registration, which is why no setup is asked of you.
  • More than one organization. Every tool takes an optional organization name. Without one, Valet uses the organization you joined first.

What it can do

  • publish_site — create an artifact or replace its contents. Works signed in or not.
  • list_sites — your organization's artifacts. Signed in only.
  • get_site — one artifact's URL, who can see it, and when it expires.
  • set_site_access — public or private. Signed in only.
  • delete_site — take an artifact down.

Password-protected access, folder uploads from disk, and everything to do with agents stay with the CLI. The tool list grows at this one address rather than at a second connector, so a client that has connected once picks up what comes next.

Not the same as an agent's MCP server

An agent you build can expose its own tools over MCP, at its own endpoint — that is an MCP channel, and calling it spends an agent turn. This server is the Valet platform itself, and its tools are ordinary API calls that return in under a second. The two hold separate credentials: a token for one is refused by the other.