claude-task-manager (ctm) — a single-writer Kanban board for coordinating Claude Code agents

[GeRiY]

claude-task-manager (ctm) — a shared Kanban board for orchestrating Claude Code agents

Hi all 👋

I'd like to share claude-task-manager
(ctm) — a standalone, dockerized, multi-project Kanban task manager built
specifically for coordinating Claude Code agents (a main agent plus teammate
subagents). It fits the Multi-Agent Systems / orchestration & coordination bucket.

The problem it solves: once you run more than one agent — or just a long session — you lose
track of what's done, what's blocked, who's working on what, and what the original goal
even was. ctm gives the agents (and you) a persistent, shared source of truth without
burning tokens re-reading state.

What it is

• A token-efficient CLI (task.sh, jq-based) that is the single, atomic writer for
  the board. Agents call compact commands (summary, list, add, status, assign,
  dep, …) instead of re-reading/rewriting a big JSON blob every turn.
• A shared browser board (Docker + PHP) that visualizes todo → in_progress → review → done across every registered project — the human drags cards, the agents move them, and
  both go through the same writer (the browser never touches the JSON directly).
• A one-command installer (ctm init) that drops a project-local skill (SKILL.md +
  a path-baked task.sh wrapper), a set of generic teammate agents
  (ctm-frontend-developer, ctm-backend-developer, ctm-code-investigator), and the
  hooks that auto-allow task.sh and inject an inbox notification after each call.

Engineering decisions worth noting

• Single-writer invariant. Every mutation — from agents and from the board UI — goes
  through task.sh, with an atomic lock, consistent timestamps, history, and an
  events.jsonl-based inbox. No corrupted/duplicated state, no "the agent and I both wrote
  the file" races.
• No Docker for agent work. Only the board + write API run in a container; the
  task.sh calls agents make are plain host bash with TM_DIR baked into the wrapper, so
  they run standalone (no cd, no permission prompt).
• Coordinator/worker split. The main agent doesn't implement — it captures requests
  as tasks, breaks them down, assigns them to teammates, reviews, and drives the
  review → done | reassign handoff loop. Teammates claim, work, and report under their
  own --as <name> identity.
• Durable project context. Beyond tasks, a context.json (goal, decisions,
  constraints, current focus) lets a fresh session resume without re-interviewing the
  user.
• Security-minded board. The Docker port binds to 127.0.0.1 only; the write endpoint
  runs an explicit command allowlist (no rm/raw/archive exposed to the browser); every
  request is checked against the registered project list.
• Bilingual (EN/HU) UI, with the board's language surfaced to the agent via a small
  stderr reminder — never stored in the task data.

Links

• Repo: https://github.com/GeRiY/claude-task-manager
• npm: npm install -g @mgeri1993/claude-task-manager — package
• How-to (user's perspective, "what to type" to Claude Code):
  English ·
  Magyar
• License: MIT

Feedback, questions, and setup ideas very welcome — and if it looks like a fit for the
list, I'd be glad to open a proper submission PR following CONTRIBUTING.md. Thanks for
maintaining such a useful collection! 🙏

[karimi1385ka-arch]

vary good