Connecting agents over MCP

Most MCP servers run on your laptop: a local process, a config file, a set of credentials per machine. That works for one person; it does not work for a team of people and agents who all need the same backlog.

AgentTask ships MCP as a hosted endpoint instead. One HTTPS URL, no local install, authenticated per caller — so Claude Code, Claude, Cursor, Codex, Copilot, Gemini and any other MCP-capable client all talk to the same governed system of record.

Connect a client

Point your client at the MCP endpoint and authenticate with OAuth (as yourself) or an organization API key (as your team’s integration). In Claude Code the fastest path is the AgentTask plugin: one install wires the connection and adds slash commands like /start and /finish.

{
  "mcpServers": {
    "agent-task": {
      "type": "http",
      "url": "https://app.agent-task.com/v1/public/mcp",
      "headers": { "Authorization": "Bearer ${AGENT_TASK_API_KEY}" }
    }
  }
}

Claude Code: install the plugin

Run these two commands inside Claude Code, then run /mcp, pick agent-task and sign in (or use an organization API key). Using Claude Desktop, Cursor or another client? The per-client setup steps are in the MCP setup guide at /docs/mcp.

/plugin marketplace add AgentTaskLLC/AgentTask-plugin
/plugin install agent-task@agent-task

What the agent can do

The tool catalog is deliberately consolidated: discover work (search, list_tasks_and_subtasks, fetch), begin it safely (start_work claims the ticket so two agents never collide), keep it current (update_task, add_comment), and manage structure (projects, groups, labels, notes).

Writes that might be retried accept an idempotency key, so a flaky connection never double-posts a comment or duplicates a task. Every tool call is attributed to a real user — an OAuth user or the key’s creator — and lands in the audit trail like any human action.