New: watch your coding agents work, step by step, on a live board. Try the demo, no account

A shared task list for Claude Code

By Jake Marsh, who builds Taskhold. Updated 2026-10-31. Facts about other tools checked 2026-10-26. 8 min read.

Short answer: Taskhold is a hosted task list Claude Code reads and updates over MCP. Add it with one claude mcp add command or by pasting a setup URL. You queue tasks in a browser; Claude Code plans each one into steps, reports progress you watch on a live board, asks blocking questions there, and finishes with a summary.

Claude Code's own task list lives in a folder on your machine and shows in the terminal. On recent models it stays empty unless you turn the task tools back on, and you can't see it from your phone or from Cursor.

Taskhold is a hosted task list that you and Claude Code both use. You add and reorder tasks in the web app. Claude Code reads the same list through MCP tools, breaks a task into steps, reports progress you watch on the live board, blocks with a question when it needs you, and completes it with a summary. Every change is marked as made by you or by an agent. To also see the running step inside the terminal, one command sets up the status line (taskhold-mcp statusline --install, after the one-line install), or use the status line script or the status line generator.

The live board, drawn with the same progress bar component the app uses. A working task with five steps, and a blocked task waiting on a reply.

Setup in one paste

The quickest route is a setup URL. Make one in Taskhold Settings (it expires after 3 days), paste it into Claude Code and say "set this up". Claude Code fetches the markdown, which carries a fresh key, the claude mcp add command below and the CLAUDE.md block, then calls taskhold_list so Settings flips to Connected. The quickstart walks through the same first run.

Or connect with no key to copy: add the URL, then run /mcp in Claude Code, pick taskhold and choose Authenticate. A browser window opens. If you're signed in to Taskhold you click Allow; if not, it connects right away, and the first time your agent uses Taskhold it gives you a link to start the 7-day trial or use your existing account.

claude mcp add --transport http --scope user taskhold https://taskhold.com/api/mcp

If you'd rather add it yourself, this is the command the setup file gives Claude Code. It registers a remote (Streamable HTTP) server for your user, so every project sees it:

claude mcp add --transport http --scope user taskhold https://taskhold.com/api/mcp --header "Authorization: Bearer <your atd_sk_ key>"
Run in a terminal. Check with claude mcp list. Docs: code.claude.com/docs/en/mcp

Permission prompts on every progress report get old fast. An allow rule of the form mcp__<server>__* covers every tool of one server and nothing else, so this in ~/.claude/settings.json lets the taskhold calls through while Claude Code keeps asking about edits and shell commands:

{
  "permissions": {
    "allow": ["mcp__taskhold__*"]
  }
}
Docs: code.claude.com/docs/en/permissions (MCP tool rules)

Or install the Claude Code plugin

The plugin bundles the server config and the Taskhold agent skill, which teaches Claude Code the claim, note, block and complete workflow. Run these inside Claude Code. The plugin connects with OAuth: the first time, run /mcp, pick taskhold and sign in (no key to paste):

/plugin marketplace add gettaskhold/taskhold-plugin
/plugin install taskhold@taskhold
Source: github.com/gettaskhold/taskhold-plugin (skill and config only, MIT).

Any agent that reads Agent Skills can add the skill on its own. You still connect the server with the command above:

npx skills add gettaskhold/taskhold-plugin

Optional: status line and terminal board

The Taskhold CLI puts the task in progress, its bar and the running step in Claude Code's status line, and shows the live board in a terminal. It comes from taskhold.com and installs in ~/.taskhold, no sudo:

curl -fsSL https://taskhold.com/install.sh | sh
taskhold-mcp login
taskhold-mcp statusline --install
taskhold-mcp watch
Needs Node.js 18+ or Bun. Updates: taskhold-mcp update. Details, checksums and uninstall: taskhold.com/install

Tell Claude Code how to use the list

Claude Code reads CLAUDE.md at the start of every session: ./CLAUDE.md or ./.claude/CLAUDE.md for one repo, ~/.claude/CLAUDE.md for all of them. Paste this block into one of those. The setup file carries the same text:

## Tasks (Taskhold)

The person and I share a task list through the taskhold MCP tools. They watch a live board while I work.
Pass project: "<repo folder name>" on every call (created on first use, no ids needed).
- Session start: taskhold_get_current and taskhold_updates, both with project. Keep the cursor.
- New request: taskhold_begin with title, project, steps and agent: "Claude Code in <repo>". It files the task, plans the steps and starts the first one. Queued work: taskhold_next with project and agent takes the next To Do task and starts it; after each one, call it again until it says To Do is empty.
- Project notes in a reply (stack, commands, what done means, what to avoid) are the person's rules for every task in that project.
- Each step done: taskhold_complete with start_next: true. Steps drive the progress bar.
- Long step: taskhold_progress on the step (N of M, or a percent) with a short message. Every few minutes at most.
- Need a decision: taskhold_block with one specific question, then taskhold_updates with task_id, since and wait: 50 until the reply arrives.
- Done: taskhold_complete on the task with what changed and how I checked it.
- A TODO or bug I'm not fixing now: taskhold_create with project (or taskhold_create_from_todo with file and line).

Repos set up for other agents often keep their instructions in AGENTS.md. Claude Code v2.1.277 and later reads that file only when there is no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md in the working directory or above it. Once you add a CLAUDE.md, put a line with @AGENTS.md in it so both load, and keep the Taskhold block in just one of them. Your ~/.claude/CLAUDE.md doesn't count for that check, so a user-level block loads either way.

With that in place, starting a session with "work on the next task" is enough. Claude Code checks for a task in progress first, then takes the highest-priority todo.

Watch Claude Code work on the live board

Claude Code claims a task, plans its steps, reports as it goes, blocks with a question and finishes with a summary, all through taskhold_* calls; the MCP tools reference has each one with its arguments. You watch the result on the live board from a browser or your phone. The no-account demo shows a run, and how to monitor a coding agent's progress explains the status dot and time estimate.

Claude Code passes a label when it starts a task, such as "Claude Code in api-server", so two sessions in different repos show up as two agents. Without a label, Taskhold uses the client name when the client sends one.

Claude Code has its own views too, and they pair well with this. Agent view (claude agents, a research preview) lists the background Claude Code sessions on your machine and which ones need input. Remote Control lets you drive a local session from the Claude app. The live board adds the task-level picture: steps, progress, the question an agent is waiting on, across every repo and every MCP client you connect.

Why not the built-in task list or a TODO.md

  • Claude Code's task list lives in ~/.claude/tasks/ on one machine and shows in the terminal with Ctrl+T. Since v2.1.268 it fills only on Opus 4.7, Sonnet 4.6, Haiku 4.5 and earlier models unless you set CLAUDE_CODE_ENABLE_TODO_TOOLS=1. Taskhold's list is the same on every model, machine and client.
  • A TODO.md in the repo works, but you have to open the repo to read it, and two sessions editing it at once can step on each other. Taskhold is one list you read from any browser, with history.
  • If a plain file is enough for you, keep using it. Taskhold is for when you want to queue work from anywhere, watch it move and answer questions without opening a terminal.

Removing Taskhold from Claude Code

Claude Code sends Taskhold the task text, steps, progress messages and notes it writes. Nothing comes back except the answers to its own calls, and your replies reach it only when it asks for updates.

To cut one machine off, revoke its key under Settings > API keys; its next call fails with 401. Then run claude mcp remove taskhold --scope user so Claude Code stops trying. If /mcp shows taskhold as failed or the tools never appear, Claude Code MCP server not connecting goes through the usual causes.

Share this page:XLinkedInBlueskyHacker News

Answer your coding agents from anywhere.

$12/month or $120/year after the trial. Card up front. Cancel before day 7 and you pay nothing. Every agent you run is included.

Or connect your agent first, no sign-up: add https://taskhold.com/api/mcp as an MCP server. The first time it uses Taskhold, it gives you the trial link.

Questions

Does Taskhold read my code?
No. Claude Code sends only the task text, notes, file paths and line numbers it chooses. Taskhold never connects to your repository and makes no AI model calls with your data.
Do I need to install anything?
No. Claude Code connects to the remote server at https://taskhold.com/api/mcp directly. The plugin and the CLI above are optional: the plugin adds the workflow skill, the CLI adds the status line and a terminal board.
Can several Claude Code sessions share one list?
Yes. Each session sees tasks already in progress, so a second session takes the next todo instead of the same one. On the live board each session shows under its own label.
Can I see Claude Code's progress from my phone?
Yes. The live board is a web page, so it works in a phone browser: steps, progress, the current message and any blocked question, with a reply box. Claude Code's Remote Control is the way to drive the session itself from the Claude app.
What does it cost?
$12 a month or $120 a year per person, after a 7-day trial that starts with a card. Cancel before the trial ends and nothing is charged.

Sources

Related

Updated 2026-10-31. Something wrong or out of date? Email hello@greatwork.company.