Skip to content
AgentTask

Connect Claude to your backlog over MCP

Give Claude your backlog, governed not guessed. Install the plugin, authenticate once, and Claude works your real tickets over MCP: reading tasks, posting progress, and closing them out — on the same board your team uses.

Start freeMCP setup guide

Three steps to your first agent-run ticket

1. Install the plugin

In Claude Code, add the AgentTask plugin source (a Claude Code plugin marketplace, not the crew marketplace) and install the agent-task plugin with the two commands below — it ships the MCP connection plus the /start…/finish workflow commands. Using Claude Desktop, the Claude web app or another client? The MCP setup guide has the endpoint and steps for each.

2. Authenticate

Run /mcp, pick agent-task and sign in with OAuth (the endpoint is https://app.agent-task.com/v1/public/mcp), or use an organization API key (amk_…) instead. The key is scoped to the spaces and capabilities you choose, and each request is attributed to it in the audit trail.

3. Run your first ticket

Point Claude at a ticket code such as ENG-142 (for example “/start ENG-142”). It claims the task, posts progress comments as it works, links its PR, and marks the work done — all where your team can watch.
Run inside Claude Code
/plugin marketplace add AgentTaskLLC/AgentTask-plugin
/plugin install agent-task@agent-task

Setup verified 2026-10-08 · agent-task plugin v0.3.0

Claude Desktop, the Claude web app, or Claude Code without the plugin

Claude Desktop and the Claude web app

Open Customize → Connectors → + → Add custom connector, paste the MCP endpoint below and sign in when the browser opens. On Team and Enterprise, an Owner adds it under Organization settings → Connectors, then members click Connect.

Claude Code without the plugin

Add the endpoint as a remote HTTP MCP server, then run /mcp, pick agent-task and sign in. You get the tools but not the /start…/finish commands the plugin ships.
MCP endpoint (Claude Desktop and the Claude web app)
https://app.agent-task.com/v1/public/mcp
Claude Code, no plugin
claude mcp add --transport http agent-task https://app.agent-task.com/v1/public/mcp

What the plugin adds to Claude Code

Commands are namespaced by the plugin, for example /agent-task:start. The two marked read-only never change your board; the rest can.

/agent-task:start

Map a ticket to a project and group, describe it, apply labels, claim it and set it in progress.

/agent-task:plan

Break a task into well-formed subtasks, then claim the first piece.

/agent-task:update

Sync the board with reality: capture untracked work and post status-update comments.

/agent-task:finish

Resolve subtasks, post a summary, confirm the PR is merged and set the task to done.

/agent-task:triage

Route unsorted tasks to a group and project, label them, set priority and flag duplicates.

/agent-task:organize

Reorganize tasks in a chosen scope into groups, projects and labels.

/agent-task:init

Stand up a new project through clarifying questions, with optional seed groups and tasks.

/agent-task:note

Capture a decision, meeting or spec as a note and link the related tasks.

/agent-task:standup

Your daily standup: what you moved, what is in progress and your blockers. Read-only.

/agent-task:report

A full progress report for a date range, space or project. Read-only.

/agent-task:crews

Sync your space’s crews into real Claude Code subagents and push local edits back.

/agent-task:launch

Hand a task to an agent on one of your connected machines, with a dry-run preview.

/agent-task:runs

Check agent runs: what ran, how it ended, what it delivered and what to do next.

/agent-task:schedule

Set up, list or cancel recurring or one-shot agent runs.

/agent-task:profile

Inspect agent profiles for a repository and attach one to a project.

The whole backlog, as typed tools

Create tasks

Capture work as structured tickets — title, description, project, labels — without leaving the conversation.

Update without context-switching

Post comments, change status, and record decisions on the ticket as the work happens.

Finish with a summary

Close out with a root-cause-and-fix summary and confirm the PR is merged.

Plan and decompose

Break a ticket into well-formed subtasks and claim the first piece.

Crew personas

Engage a saved crew (an AI persona) as a Claude sub-agent for reviews and specialised work.

Sync to sub-agents

Pull your space’s crews into real Claude Code sub-agents and skills, and push local edits back.

The difference on a real backlog

The difference on a real backlog
FeatureWithout AgentTaskWith AgentTask
Where work livesScattered across chat threadsOne shared board — people + agents
Progress visibilityLost when the session endsLive comments + status on the ticket
AccountabilityNo record of who did whatEvery MCP request on the audit trail (request level)
HandoffRe-explain context each timeTicket carries the plan, PR, and history

Bring Claude to your board

Sign up and wire up your first agent-run ticket in minutes — or talk to us about a rollout.

Start freeTalk to sales

Wondering how this compares? Claude vs alternatives.

Frequently asked questions

Three steps: install the agent-task plugin in Claude Code (it ships the MCP connection plus the /start…/finish workflow commands), authenticate once with OAuth or an organization API key (amk_…), then point Claude at a ticket. Claude claims the task, posts progress comments as it works, links its PR, and marks the work done — all where your team can watch.

The whole backlog is exposed as typed tools: Claude can create tasks, post comments and status changes, break a ticket into subtasks and claim the first piece, engage saved crew personas as sub-agents, and close work out with a summary. Its API and MCP requests are attributed to its credential on the audit trail.

No — MCP and public API access are included on every plan, starting with Free (1 API key, 1 MCP connection and 2,500 combined API + MCP requests per day). The MCP connection limit is 2 on Pro, 5 on Team and 50 on Enterprise, so on Free you can connect Claude Code or Claude Desktop, not both.

No. The plugin only points Claude at the hosted MCP endpoint (https://app.agent-task.com/v1/public/mcp); it reads and writes your backlog, not your repository, and installing it shares no AgentTask source.

Auth gates everything: OAuth login or a scoped organization API key. Nothing is reachable without it, regardless of what’s installed.

Claude Code (the plugin, or claude mcp add), and the Claude web app and Claude Desktop through Customize → Connectors → Add custom connector with https://app.agent-task.com/v1/public/mcp (browser sign-in, no key needed).

The core commands talk to the hosted MCP endpoint. Two things go further, and only when you run them: /crews sync writes subagent and skill files under .claude/ in your repository, and /crews hook install adds an optional git pre-push hook (uninstall removes it). The plugin also ships an optional status-line script that you enable yourself.

No. Claude and Claude Code are trademarks of Anthropic, PBC. AgentTask is an independent product that connects to Claude over MCP; it is not affiliated with or endorsed by Anthropic.