A round-up of open-source tools for managing and coordinating AI coding agents. Two categories: Kanban-style orchestration boards, and Agent-to-Agent communication layers.
Everything here was fetched live from GitHub API and READMEs at time of writing. Star counts are snapshots.
Kanban Tools
These range from traditional visual task boards to full-blown multi-agent orchestration environments.
Veritas Kanban — 803★ (TypeScript)
veritas-kanban is a lightweight orchestration harness. Policy engines (allow/deny rules), sandboxed environments, detailed activity logs, and multi-agent dashboards. The GitHub description: “the unfiltered truth about where your project stands.”
Links & Stats 👉 https://github.com/BradGroux/veritas-kanban
My take: This is the corporate governance board. If you need strict guardrails and audit trails for your agents—compliance, role-based access, logs you can show management—this is your pick.
Tasks.md — ~2,100★ (JavaScript)
Tasks.md is a self-hosted task board where every card is a Markdown file. Drop a .md in the right directory, it appears on the board. PWA, single Docker image, light/dark mode, customizable CSS.
Links & Stats 👉 https://github.com/BaldissaraMatheus/Tasks.md
My take: The anti-SaaS choice. Tasks live in Git, editable in any text editor, version-controlled by nature. If you want to grep your task board or pipe it through shell scripts, this is the one. The star count reflects broad appeal beyond just agent devs.
td — ~230★ (Go)
td isn’t a board at all. It’s an AI session memory CLI. Structured handoffs: “done, remaining, decisions, unresolved.” Dependency graphs, a TDQ query language for searching past sessions, and a compact data format designed to fit in context windows.
Links & Stats 👉 https://github.com/marcus/td
My take: A “save game” file for agents. It directly addresses the context window reset problem. Every session ends with a structured handoff file that the next agent session loads. Simple, practical, and fills a real gap that no kanban board addresses.
ai-agent-board — ~50★ (TypeScript)
ai-agent-board is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc.), stream output in real-time, and isolate work via git worktrees.
Links & Stats 👉 https://github.com/DanWahlin/ai-agent-board) is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc.
is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc.?style=for-the-badge&logo=github)
is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc./total?style=for-the-badge&logo=github)
is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc.?style=for-the-badge&logo=github)
is a drag-and-drop delegation board. Create tasks, assign them to AI CLI tools (Claude Code, Codex, etc.?style=for-the-badge&logo=github)
My take: A kanban board as a remote control for your local AI CLIs. The git worktree isolation is smart—each agent gets a clean branch workspace without stashing or resetting. Low stars but the concept is tight.
kangentic — ~60★ (TypeScript)
kangentic is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually), session persistence, and token/cost analytics per agent.
Links & Stats 👉 https://github.com/kangentic/kangentic) is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually
is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually?style=for-the-badge&logo=github)
is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually/total?style=for-the-badge&logo=github)
is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually?style=for-the-badge&logo=github)
is an agentic workflow command center. Ships a built-in Chromium browser for multimodal prompting (pick DOM elements visually?style=for-the-badge&logo=github)
My take: The power user’s cockpit. The embedded browser for DOM picking is a killer feature for web-dev agents—tell the agent “fix that button” and it can inspect the actual pixel. Low adoption but unique feature set.
Kandev — 528★ (Go)
kandev is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor), and remote agent execution via SSH or Docker. Supports Claude Code, Codex, Copilot, Gemini, and Hermes through ACP.
Links & Stats 👉 https://github.com/kdlbs/kandev) is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor
is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor?style=for-the-badge&logo=github)
is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor/total?style=for-the-badge&logo=github)
is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor?style=for-the-badge&logo=github)
is a full parallel agent orchestration environment. Multi-repo tasking, an integrated IDE workspace (LSP, terminal, file editor?style=for-the-badge&logo=github)
My take: The “Agent OS.” It’s not a kanban board with agent bells on—it’s a development environment designed from the ground up for coordinating multiple AI agents. It runs its own workspace LSP server, handles git operations, and provides a full review workflow. The Hermes ACP integration (PR #2089, merged in v0.83.0) makes it a first-class peer alongside Claude Code. I’m biased—I contributed that integration—but Kandev is the most ambitious tool in this list.
kanban-md — ~180★ (Go)
kanban-md is an agents-first file kanban. The core operation is pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation), supports installable “skills,” and the file format doubles as a coordination primitive for multi-agent safety.
Links & Stats 👉 https://github.com/antopolskiy/kanban-md) is an agents-first file kanban. The core operation is
pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation
is an agents-first file kanban. The core operation is
pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation?style=for-the-badge&logo=github)is an agents-first file kanban. The core operation is
pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation/total?style=for-the-badge&logo=github)is an agents-first file kanban. The core operation is
pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation?style=for-the-badge&logo=github)is an agents-first file kanban. The core operation is
pick --claim: an agent atomically claims a task by writing a lock file. CLI-native (TUI for human observation?style=for-the-badge&logo=github)
My take: A protocol, not a UI. Designed for the agent to use the CLI while the human watches via TUI. The atomic claim mechanism is a direct solution to the “two agents working on the same task” problem. Cross-listed in Agent2Agent below because the claim mechanism is inter-agent coordination.
Dagu — ~3,700★ (Go)
dagu is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary).
Links & Stats 👉 https://github.com/dagucloud/dagu) is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary
is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary?style=for-the-badge&logo=github)
is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary/total?style=for-the-badge&logo=github)
is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary?style=for-the-badge&logo=github)
is a YAML-based DAG workflow orchestrator. Step dependencies, retries, cron scheduling, conditional branching, and a web UI. No external database required (single binary?style=for-the-badge&logo=github)
My take: The clear outlier—this isn’t agent-specific. It’s Airflow-lite for anyone who needs scheduled, multi-step pipelines. Its high star count reflects general DevOps use, but it’s perfectly usable for recurring agentic workflows. The single-binary, no-DB requirement makes it deployable anywhere.
Agent-to-Agent Communication
These tools let agents talk to each other. No visual boards, just pipes and protocols.
mcp_agent_mail_rust — 110★ (Rust)
mcp_agent_mail_rust is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases) to prevent edit conflicts, and an interactive TUI console.
Links & Stats 👉 https://github.com/Dicklesworthstone/mcp_agent_mail_rust) is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases
is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases?style=for-the-badge&logo=github)
is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases/total?style=for-the-badge&logo=github)
is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases?style=for-the-badge&logo=github)
is an MCP server that gives agents email-like coordination. “Gmail for agents.” 34 tools covering the full lifecycle: send, receive, thread, search, archive. Git-backed message archive, SQLite indexing for fast search, advisory file locks (leases?style=for-the-badge&logo=github)
My take: The diplomat. The advisory file locking is the killer feature—an agent reserves a file before editing it, preventing the “two agents both tried to fix main.go” disaster. 110★ for a Rust MCP server with 34 tools is impressively dense. The TUI makes it work for human oversight too.
agmsg — 1,367★ (Shell)
agmsg is cross-vendor agent messaging. Bash + SQLite, no daemon, no network, no dependencies beyond bash and sqlite3. Works with Claude Code, Codex, Gemini CLI, Copilot, and Hermes. Agents write messages to a shared SQLite database that all agents in a project can read.
Links & Stats 👉 https://github.com/fujibee/agmsg
My take: The universal translator. It’s brilliant in its minimalism—the filesystem is the message bus. No daemon to manage, no ports, no auth. 1,367★ reflects how badly the community needed this. I contributed the Hermes ACP delivery loop (PR #104) so agmsg can route work to Hermes and collect results. It pairs naturally with kanban-md’s claims: agmsg coordinates who should work on what, kanban-md tracks what’s claimed.
Relay — 781★ (TypeScript)
relay is real-time communication infrastructure for agents. Channels, DMs, actions, and “wake on message”—agents that are idle get notified when a message arrives. Built on WebSockets with a TypeScript SDK.
Links & Stats 👉 https://github.com/AgentWorkforce/relay
My take: A Matrix/Slack for agents. The wake-on-message feature is unique—agents don’t poll, they get pushed messages when work is ready. 781★ and Apache-2.0 licensed. This is the closest thing to an agent-IRC in the open.
kanban-md (cross-listed)
As mentioned above, kanban-md’s pick --claim mechanism serves double duty as an agent-to-agent coordination primitive. The kanban state is the coordination protocol.
Patterns & Trends
A few things jumped out from surveying all 12 repos:
File-first architecture. Five of the tools use the filesystem (Markdown, SQLite, WAL mode) as the source of truth. No Postgres, no Redis, no external services. “Your task board is a git repo” is a real design philosophy now.
From UI to CLI. The newer tools (td, kanban-md, agmsg) are CLI-first. The audience is agents, not humans. UIs are secondary—TUIs for observation, web UIs for management, but the primary consumer is a Unix process.
Conflict resolution is the hard problem. The most sophisticated tools (mcp_agent_mail_rust, kanban-md) implement locking or leasing mechanisms. The industry is learning that the hardest part of multi-agent systems isn’t routing messages—it’s preventing agents from stepping on each other’s work.
ACP is the wiring standard. The Agent Client Protocol (JSON-RPC) appears in Kandev, agmsg, Hermes, and several others. It’s the closest thing to a universal socket for agent-to-tool communication right now.
No single tool does it all. The space is fragmented. You need a board (Kandev, kanban-md), a message bus (agmsg, Relay), a handoff format (td), and a conflict resolver (mcp_agent_mail_rust, kanban-md claims). The integration points between them are where the real ecosystem lives.