Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@DooistAdd a task to review the API documentation by Friday with priority 1 @work"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Dooist
The agent todo app. Task management built for AI agents, not people.
Why
Every Claude Code session starts blank. No memory. No state. Your agent doesn't know what it was working on yesterday.
So you re-explain. Re-paste the spec. Re-describe the goal. Every single time.
That's not a workflow. That's babysitting.
The Fix
Dooist is an MCP server where every task carries rich context — intent, spec plan, desired outputs, current status. Any Claude session picks up any task and knows exactly what to do.
No claude --resume. No re-explaining. Context lives in the task, not the session.
Next session:
Agent-Native Design
Most developer tools with "AI support" are human tools with an API bolted on. That's backwards.
Agents don't need pretty UIs. They need structured data, rich context fields, and tool interfaces that match how they actually work — stateless, parallel, fast.
list_tasksshows titles and a[has context]flag. Scannable. No noise.get_taskreturns the full markdown body — intent, plan, desired outputs.create_tasksupports parallel calls. Create 10 tasks in one message.
Information architecture matters for agents too. Don't dump everything at once.
What This Looks Like
Run headless Claude Code against your inbox. It triages emails into tasks with full context — who sent it, what they need, deadline. Next Claude session picks up those tasks and executes.
Slack messages, meeting transcripts, texts — all of it generates work. An agent parses your inputs, creates tasks with full context, prioritizes them, and starts working the queue.
No human in the loop for task creation. The agent manages its own backlog.
Install
Restart Claude Code. Done.
Teach Claude Your Workflow (Optional)
Installing the MCP server gives Claude the tools. This step teaches it when to use them — proactively suggesting tasks, checking your backlog, adding rich context.
Dooist ships a DOOIST_CLAUDE.md file you can import into your CLAUDE.md. One line:
That's it. Every Claude Code session now knows how to use Dooist — checking tasks on planning conversations, suggesting task creation when you mention future work, and including rich context automatically.
What this changes:
"We should fix that later" → Claude asks "Want me to track that in Dooist?"
"What should I work on?" → Claude checks your task list
Bugs found during work → Claude suggests creating a task
Session ends mid-task → Claude updates context so the next session picks up
To remove: Delete the @~/.claude/rules/dooist.md line from your CLAUDE.md.
Usage
Just talk:
Tools
Tool | What it does |
| Add task with context, due date, priority, labels, parentId |
| Full task details including rich context body |
| Filter by project, label, status, parentId — scannable summaries |
| Mark done |
| Change anything including context and parent |
| Remove (cascades to subtasks) |
| Append timestamped status update to a task |
| Due today + overdue |
| Next N days |
| Find TODO/FIXME/HACK comments in codebase |
| All projects |
| New project |
| All labels |
| New label |
Features
Rich task context: Markdown body with intent, plan, outputs, status
Subtasks: Nest tasks with
parentId, cascade delete, filter by parentTask updates: Timestamped activity log per task
Natural language dates: "tomorrow", "next monday", "in 3 days"
Priorities: P1 (urgent) through P4 (low)
Labels: @work, @personal, @urgent
Projects: Group related tasks
Truncated IDs: Use first 8 chars of any task ID
Persistent: SQLite at
~/.dooist/Self-hosted: Your data stays on your machine
Custom Database Location
Development
License
MIT