I spent some time digging into the current landscape of tools for two related problems: visual task management for AI agents (kanban boards that understand agents) and agent-to-agent communication (how agents talk to each other).
Some of these I already run day-to-day. Others are new to me. Here’s what I found.
Kanban
veritas-kanban — ★803, TypeScript
Links & Stats 👉 https://github.com/BradGroux/veritas-kanban
An AI agent orchestration harness with a kanban UI. Ships policy engines (allow/deny rules for what agents can do), sandboxed execution environments, detailed activity logs, and multi-agent dashboards.
Vibe: strict governance. If you need audit trails for every agent action and want to enforce guardrails per-task, this is your board. The README pitches it as “the unfiltered truth about where your project stands” — the governance angle is front and center.
Tasks.md — Self-hosted, Markdown-based
Links & Stats 👉 https://github.com/BaldissaraMatheus/Tasks.md
A web UI where cards ARE markdown files. Create lanes, tags, cards — all stored as .md files on disk. Ships as a single Docker image.
Vibe: simplicity-first. No database, no API to learn — your task board is just a folder full of markdown files. Version control your kanban board alongside your code. Light/dark mode, responsive.
td — ★235, Go
Links & Stats 👉 https://github.com/marcus/td
Minimalist CLI for tracking tasks across AI coding sessions. When your context window ends, your task state doesn’t disappear — it persists in a simple Go binary.
Vibe: terminal-native, zero fuss. No UI, no web server, no Docker. Just td add, td list, td done. Built specifically for the agent-in-a-terminal workflow.
ai-agent-board — ★48, TypeScript
Links & Stats 👉 https://github.com/DanWahlin/ai-agent-board
Drag-and-drop kanban board that delegates tasks to AI agents — GitHub Copilot, Claude Code, OpenAI Codex, and OpenCode. Real-time streaming output, task groups, git worktree isolation per task.
Vibe: the most agent-aware board in this list. It doesn’t just visualize tasks — it runs them. You drag a card to “In Progress” and an actual agent starts working in its own git worktree. Streaming output means you watch the agent code in real time.
kangentic — ★66, TypeScript
Links & Stats 👉 https://github.com/kangentic/kangentic
Desktop kanban board for orchestrating agentic workflows. Electron app with a focus on workflow orchestration through a kanban interface.
Vibe: polished desktop UX. Still early (66★) but the concept is right — a native-feeling board for agents. Electron, so not terminal-native, but the workflow-angle is interesting.
kandev — Multi-agent orchestration dashboard
Links & Stats 👉 https://github.com/kdlbs/kandev
Full web dashboard where you create tasks, assign them to different agents (Claude Code, Codex, Copilot, Gemini, Hermes), review output, and manage PRs. ACP-native — Hermes already ships a kandev ACP integration (merged PR #2089).
Vibe: the heavy lifter. Most feature-complete multi-agent kanban I’ve worked with. Real ACP protocol integration, not a shell wrapper. If you’re running multiple agent types and need a command center, this is the one.
kanban-md — ★179, Go
Links & Stats 👉 https://github.com/antopolskiy/kanban-md
File-based kanban board for autonomous agentic loops. CLI + TUI, no daemon, no database. Cards are markdown files. Ships its own “agent skills” — predefined workflows agents can execute from the board.
Vibe: terminal-native and agent-first. The same “markdown as state” philosophy as Tasks.md, but built explicitly for agent-in-the-loop workflows. Skills system means agents can self-assign from the board.
dagu — ★3,691, Go
Links & Stats 👉 https://github.com/dagucloud/dagu
Self-hosted workflow orchestration in one binary. YAML DAG definitions, cron scheduling, AI agent steps, web UI. Lightweight Airflow alternative — no database, no Python dependencies.
Vibe: workflow engine, not a kanban board — but the overlap is real. DAGs, cron, agent steps, human-in-the-loop gates. 3,600★ tells you it’s solving a real problem. If your “kanban” is really “I need to chain agent tasks in a DAG,” dagu is the dark horse here.
Agent-to-Agent Communication
mcp_agent_mail_rust — ★110, Rust
Links & Stats 👉 https://github.com/Dicklesworthstone/mcp_agent_mail_rust
Rust MCP server for multi-agent coordination: 34 tools, Git-backed archive, SQLite indexing, advisory file locks, interactive TUI console. Implements a mailbox metaphor — agents send messages to each other’s inboxes.
Vibe: serious engineering. 34 MCP tools is a LOT. Git-backed archive means every message is version-controlled. The advisory file locking mechanism lets agents coordinate access to shared resources (files, APIs) — solving the “two agents writing the same file” problem.
agmsg — ★1,367, Shell + SQLite
Links & Stats 👉 https://github.com/fujibee/agmsg
Cross-vendor messaging for CLI AI coding agents. Let Claude Code, Codex, Gemini CLI, and Copilot talk to each other in one team. Bash + SQLite — no daemon, no framework.
Vibe: lightweight and effective. 1,300★ tells the story. I use this one — it’s a single shell script and a SQLite database. No MCP server to run, no daemon to manage. Agents write to the same SQLite DB, poll for messages, and coordinate. Bash means it works everywhere.
relay — ★781, TypeScript
Links & Stats 👉 https://github.com/AgentWorkforce/relay
Real-time communication for agents. Wake on message, channels, DMs, actions. Designed for orchestrating agent teams with a Slack-like channel model.
Vibe: structured messaging with sleep/wake. Agents register on channels, go to sleep, and get woken up when a message arrives in their channel. Channel system gives you routing — agent A posts to “code-review” channel, agent B picks it up. Actions extend beyond text (trigger functions, run workflows).
kanban-md (also listed)
kanban-md appears in both categories because its file-based model doubles as a coordination primitive — agents read and write the same markdown board files to communicate task state. It’s not real-time messaging, but for many workflows, “agent puts a card in the ’needs review’ lane” IS the message.
How They Stack Up
Different layers of the stack. The kanban tools range from pure visualization (Tasks.md, td) to full agent dispatch (ai-agent-board, kandev). The A2A tools cover message passing (agmsg, mcp_agent_mail_rust) to channel-based orchestration (relay).
The combination that’s most interesting to me:
- kandev for task management and agent dispatch
- agmsg for lightweight inter-agent messaging
- relay or mcp_agent_mail_rust for structured, production-grade coordination
No single tool covers all three. The ACP protocol is starting to bridge them, but the space is still early — 6 months from now this list will look very different.