CLI

Install the open-source HiveFlow CLI: a router for Claude Code, Codex, Gemini and Aider, a native agent with its own tool-use loop, and workers that turn kanban cards into pull requests.

The HiveFlow CLI is the terminal door into your workspace. It is MIT-licensed and fully auditable at github.com/hiveflowai/hiveflow-cli — which matters, because the recommended install is a curl | bash one-liner that touches your machine.

It has three layers, and you can use any one of them on its own.

Layer	What it does
Router	Infers the task type and launches the best AI coding CLI you already have installed
Native agent	Its own agentic loop — reads, edits, runs commands, searches, delegates to subagents — with no external CLI required
Workers	Local agents that watch your HiveFlow kanban boards and work the cards, with human-in-the-loop gates

Install

One line, no npm or git required:

curl -fsSL https://hiveflow.ai/install.sh | bash

With npm, from the @hiveflow/cli package:

npm install -g @hiveflow/cli

From the repo, for development:

git clone https://github.com/hiveflowai/hiveflow-cli.git
cd hiveflow-cli && ./install.sh

The installer links the binary into ~/.local/bin (or /usr/local/bin) and offers to install any AI coding CLIs you're missing. Four names run the same program: hiveflow, hive, hf and flow.

First run

hiveflow

  🐝 Welcome to Hiveflow 👋

hiveflow (auto·auto) ❯ fix the login bug
  › Router: task 'quick-fix' → codex

hiveflow (auto·auto) ❯ /agent refactor src/auth.js and run the tests
hiveflow (auto·auto) ❯ /worker list
hiveflow (auto·auto) ❯ /help

Sign in with your HiveFlow account the first time — the same account as the web app. /help lists every command, and /help <command> explains one.

The router

Type a task in plain language and the router classifies it (quick fix, refactor, research, review…) and hands it to the installed CLI that is strongest at that shape of work — Claude Code, Codex, Gemini or Aider. You keep your existing subscriptions and tools; the CLI just picks between them so you don't have to.

The native agent

/agent <task> runs HiveFlow's own tool-use loop against the Claude, ChatGPT or Gemini API, with no dependency on any external CLI. It reads and edits files, executes commands, searches the codebase and delegates to subagents, and it supports permissions, sessions, hooks, skills and MCP servers.

Workers

/worker connects local agents to the kanban boards in your workspace. A worker watches a column, picks up cards, does the work and moves the card along — a support ticket becomes a pull request, a sales playbook runs itself — with metrics, evals and a CI gate, and human approval wherever you place a gate.

The same workers show up in the mobile apps and over MCP, so you can check on them from your phone or ask Claude "why isn't the onboarding card being picked up?".

MCP from the terminal

The CLI speaks MCP, and HiveFlow itself is available as a remote MCP server. If you'd rather wire HiveFlow into Claude Code or Codex directly instead of through this CLI, see HiveFlow in Claude and Claude Code — the endpoint is https://mcp.hiveflow.ai with OAuth 2.1.

Desktop & mobile — the Electron app and the phone and watch apps
HiveFlow in Claude and Claude Code — MCP connector and Claude Code plugin
API reference — the same operations over HTTP

CLI