serpdive-cli

One command to give every AI coding agent on your machine live web search.

npx -y serpdive-cli init --key sd_live_YOUR_KEY

Free key, no card: serpdive.com/dashboard/keys.

What it wires

The MCP server — your agent gains a serpdive_search tool. One call returns the extracted, answer-ready text of live pages, not a list of links to go fetch. Hosted, so there is nothing to install or keep up to date.

A skill — the half people forget. The tool lets an agent search; the skill tells it which model to pick, how to keep the response small, and what the error codes mean. Without it the agent guesses.

Claude Code · Codex · Cursor · Windsurf · VS Code · Zed · OpenCode · Gemini CLI · Cline · Goose · Grok Build · GitHub Copilot CLI · Claude Desktop · Antigravity — whichever are installed. The rest are skipped.

Related MCP server: Deep Research MCP Server

Options

--key <sd_live_…>   your API key. Defaults to $SERPDIVE_API_KEY.
--agent <id>        one agent only (repeatable): codex, cursor, claude-code, …
--local             run the server yourself via npx, instead of the hosted one
--skip-skill        register the tool but install no skill
--skip-mcp          install the skill only
--dry-run           print what would run, change nothing

How it writes your configs

It doesn't. MCP registration is handed to add-mcp, which exists to know each agent's config format and is downloaded ~225k times a week.

That is not laziness, it is the fix for a real bug. The first version of this CLI wrote the configs itself, assuming every agent used the same { "mcpServers": … } JSON. They do not: Codex keeps its servers in TOML as [mcp_servers.name] with a type = "http" and an http_headers table, OpenCode uses a different root key, VS Code another, Zed another. Agents were reported as wired while receiving nothing — a green line for a lie. Sixteen formats is somebody's full-time job, and somebody is already doing it.

What is still ours:

• Global configs only. A CLI run from some directory has no business adding a server to that project's committed files.

• The key travels in a header, not in the URL, so it stays out of shell history and screen recordings. --local puts it in the environment instead.

• The skill is a plain file copy — no config parsed, nothing of yours at risk. Running twice changes nothing the second time.

Nothing phones home.

Without the CLI

The hosted server needs no install at all:

claude mcp add --transport http serpdive https://mcp.serpdive.com \
  --header "Authorization: Bearer sd_live_YOUR_KEY"

Docs: serpdive.com/docs · MIT