@gavai/builder-mcp) is the hand-curated MCP surface for the operations a page editor actually performs — reading workspaces, patching the page document, publishing versions, and walking the block and component schemas. It is bundled with the gavAI CLI and reuses your CLI session, so once you have run gavai login, registering it with Claude is a one-line config change. By the end of this page you will know what tools the server exposes, how to wire it into Claude Desktop, and what an end-to-end edit looks like on the wire.
At a glance
| Package | @gavai/builder-mcp (bundled in the gavAI CLI) |
| Transport | stdio (JSON-RPC) |
| Command | gavai mcp |
| Auth | ~/.config/gavai/credentials.json, written by gavai login |
| Tools | 14 (see below) |
Install and authenticate
The server ships inside the gavAI CLI, so installing the CLI is installing the server. See CLI overview for the install command on your platform. After install, run the login flow once:~/.config/gavai/credentials.json. The MCP server reads that file on every tool call and refreshes the token automatically — you do not handle bearer tokens by hand.
Register the server
- Via CLI (recommended)
- Manual config
The CLI knows where Claude Desktop’s config lives on each OS and can write the entry for you:This adds a
gavai-builder server entry pointing to gavai mcp and picks up the correct absolute binary path. Restart Claude Desktop to load the server.Tools
The server registers 14 tools, grouped by what they touch.Workspace and app document
Blocks and components
Theme
Each tool dispatches to a single REST endpoint — for example,patch_tenant_app calls and publish_tenant_version calls .
How clients discover them
The names and full input schemas are returned by the standard MCPtools/list call. Claude Desktop, Cursor, and other hosts use that response to present the tool list and validate inputs before invoking them — this is part of the protocol, not a gavAI extension.
Worked example: AI adds a nav link and publishes
Here is the exchange when a user says “Add a Pricing link to the navigation on acme and ship it”: Three properties of the exchange are worth pulling out:- The AI never touches a credential. The MCP server reads
~/.config/gavai/credentials.jsonand attaches the bearer token on every API call. - Every edit is a JSON Patch.
patch_tenant_apprejects free-form rewrites; the AI must produce RFC 6902 operations against the current document. - Publishing is an explicit second call.
patch_tenant_appwrites a draft. The draft is only visible on the live site afterpublish_tenant_version. A failed publish does not roll the patch back.
Security
The server runs with the privileges of your CLI session. Any tool call the host makes — patch, publish, theme rewrite — executes against your active workspace with your credentials. Two practical rules:- Treat the connected workspace like production. AI chains tool calls; a chain ending in
publish_tenant_versionreaches your live site. Point experimentation at a test workspace. - Revoke when done. Run
gavai logoutto invalidate the session locally, or revoke it in the console under Settings → Sessions.