Integrations

Hive integrates with agent CLIs in two ways: terminal-screen detection works for the supported agents, while hooks give Claude Code and Codex precise state and conversation-ID reports.

Agent hooks

Install or remove the supported hooks with:

sh

hive integration install claude
hive integration install codex
hive integration uninstall claude

Hooks report state changes, including when an agent is waiting for approval, and let Hive resume the correct conversation after a restart. The installer is designed to be repeatable and preserves a backup of agent settings.

Without hooks, Hive reads the agent’s visible screen and uses the manifest’s resume-latest command after a restore. That can choose the wrong conversation if multiple agents share a working directory; separate them with worktrees.

Screen-detected agents

Hive includes manifests for Claude Code, Codex, OpenCode, Gemini CLI, Cursor Agent, GitHub Copilot CLI, Amp, and Aider. The latter six are detected from process and terminal-screen signals. You can inspect loaded definitions with hive agent manifests and investigate a state with hive agent explain <agent>.

To add or tune a CLI, create an agent manifest under ~/.config/hive/agent-detection/. See Custom Agent Support.

MCP server

Hive can provide agent-control tools over MCP’s standard I/O transport. Register the server with Claude Code using:

sh

claude mcp add hive -- hive mcp

The MCP tools let an agent list environments, start and prompt other agents, wait for results, read output, and operate panes. They act on the Hive session available to the hive mcp process.

Agent skill

The hive --skill option prints a set of instructions for agents that coordinate work through Hive. See Agent Skill for the recommended use.