AgentFlow

Orchestrate multiple AI agents from your terminal — k9s-style.

AgentFlow lets you run a team of AI agents (Claude, Gemini, GPT-4) in parallel, each in its own isolated tmux session, with a real-time dashboard to monitor and direct them.

Why AgentFlow?

• Local-first — no server, no cloud accounts beyond your AI provider
• Multi-provider — Claude (OAuth), Gemini, OpenAI in one workflow
• Parallel agents — each agent runs isolated in its own tmux session
• Smart routing — tasks auto-assigned by type (code/review/plan/test/security)
• Real-time TUI — k9s-style dashboard to monitor all agents at once
• Dev pipeline — automated branch → implement → review → PR flow

Installation

npm install -g agentflow-cli

Or build from source (requires Bun):

git clone https://github.com/Hitbee-dev/AgentFlow.git
cd AgentFlow && bun install && bun run build
cp dist/agentflow ~/.local/bin/agentflow

Quick Start

agentflow install          # first-time setup
agentflow auth login       # connect Claude via OAuth (browser)
agentflow agents start-all # start all agent workers
agentflow                  # open the TUI dashboard

From the TUI you can submit tasks, filter by namespace, and watch agents work in real time.

Architecture

agentflow
├── src/auth/       OAuth PKCE + OS keychain (Claude / Gemini / OpenAI)
├── src/provider/   Vercel AI SDK multi-provider abstraction
├── src/agent/      tmux sessions, heartbeat, dispatch queue, broadcast
├── src/orchestrator/ Task routing — auto by type, manual by @agent
├── src/pipeline/   Automated dev pipeline (branch → PR)
├── src/tui/        React/Ink k9s-style dashboard
└── src/cli/        Commander.js CLI commands

Runtime: Bun · UI: React/Ink · LLM: Vercel AI SDK · Auth: OS keychain

Default Agents

Name	Role	Model
planner	Feature design & specs	claude-opus-4-6
coder	Implementation	claude-sonnet-4-6
reviewer	Code quality review	claude-sonnet-4-6
security	Security audit	claude-opus-4-6
qa	Testing & verification	claude-sonnet-4-6

Key Commands

agentflow                          # TUI dashboard
agentflow run "<task>"             # submit task (auto-assigned)
agentflow run --agent coder "<task>" # assign to specific agent
agentflow chat planner "<msg>"     # one-shot streaming chat
agentflow pipeline start "<desc>"  # start automated dev pipeline
agentflow status                   # quick overview
agentflow doctor                   # health check

TUI shortcuts: ↑/↓ navigate · Enter detail · Esc back · :q quit
TUI commands: :ns <name> filter namespace · @agent <task> assign · 공지! <msg> broadcast

Supported Models

Provider	Models
Anthropic	claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5
Google	gemini-2.5-pro, gemini-2.0-flash
OpenAI	o1, gpt-4o, gpt-4o-mini

License

Apache-2.0 © Hitbee-dev