Pick an agent
Hover doesn't bundle or spawn an AI runtime. It's an MCP server that runs
inside the coding agent you already use, riding the model and subscription you
already pay for (BYO-CLI). Any MCP-capable agent works — they all get the same
grounded tools and the /mcp__hover__test_app flow.
Which agents work
| Agent | How to add | Notes |
|---|---|---|
Claude Code (claude) | claude mcp add hover -- hover-mcp | Recommended. Its allow/deny tool controls can scope a session to Hover's MCP surface. |
| Cursor | Add hover-mcp as an MCP server in Cursor's settings | Same tools + /mcp__hover__test_app. |
| Any MCP-capable agent (Codex, Gemini CLI, …) | Point the agent's MCP config at the hover-mcp command | Works as long as the agent speaks MCP. |
Install the server once (npm i -g @hover-dev/mcp), then register it in your
agent — see Install.
How Hover plugs in
Your agent is the host: it runs the model, holds the auth, and calls Hover's
tools. Hover contributes the grounded browser surface (browser_snapshot,
click_control, fill_control, crystallize_spec, …) and the workflow prompts
(test_app, guard, build, heal, optimize). It never handles your model
key — see API keys.
Because the agent owns the model, picking an agent is really picking which subscription and which permission model you want; the Hover tools are identical across all of them. Claude Code is the default recommendation for its allow/deny tool controls, which let you fence the session down to Hover's MCP surface.
The VS Code cockpit is not an agent
The optional VS Code extension is a review cockpit — a Business Map graph + a Dashboard. It drives no agent and runs no model; authoring always happens through the MCP inside your own agent.