Qalatra

Project management for the AI era.

Give your AI agent full read/write access to your tasks, daily notes, and habit log — all in one local app. Not just a task manager your agent can read. A system your agent actively operates: creating tasks, triaging your backlog, running your morning briefing, dispatching autonomous agents to complete work, and building a persistent memory of your days that makes weekly reviews and long-term planning actually useful.

Built for Claude Code via MCP. Works with any command-line AI agent.

Your data lives in a local SQLite database on your machine. Nothing goes to a server. The app is free and open source.

---

---

The idea

The AI project management space has split into two camps. One camp uses markdown files — tasks as text, everything in a folder, Claude reads the files. Simple, but no structure, no recurrence, no real query capability. The other camp keeps using traditional task managers and just asks Claude questions about them — but Claude has no write access, no persistence, no memory across sessions.

Qalatra is a third approach: a structured task database that your AI agent lives inside of.

Your agent gets full read/write access to your tasks, your daily notes, and your habit log simultaneously. That combination is what makes it different. Your tasks capture what you need to do. Your daily notes capture everything else — the meeting that went sideways, the idea you had at 2pm, the thing you wanted to remember but didn't have time to turn into a task. Together, they give your agent a real memory of your work across days and weeks.

Qalatra was designed for Claude Code and all the examples use it — but it works with any command-line AI agent. That's configurable in Settings.

In practice that looks like:

• Morning: "What's my day look like?" → your agent reads your overdue tasks, today's schedule, upcoming deadlines, and yesterday's notes, then tells you what to focus on and what to push.
• During work: "I need to follow up with Sarah about the contract, remind me Thursday" → task created, snoozed, done. No switching apps.
• End of day: "Triage me" → your agent reviews what got done, moves stale items to backlog, proposes priorities for tomorrow.
• Weekly review: "How was my week?" → your agent reads all seven daily notes plus task completion history and gives you a real account of what happened — including the small things that never became tasks but mattered.
• Agents: Assign an agent to a task. Qalatra dispatches it, captures the output, and marks the task complete. You come back to work already done.

The terminal is built in so you never leave the context. Claude runs in a panel below your task list. The daily note is one click away. Everything is in one place because context-switching is where work dies.

---

What you get

A local Electron app with four views:

View	What it shows
Priority	Today's tasks, grouped by context, drag to reorder
Project	Same tasks grouped by project
Backlog	Snoozed, unsurfaced, and future tasks
Habits	Daily habit tracker with streak display

30+ MCP tools that Claude can call in any conversation:

morning_briefing      → full briefing with priorities + overdue tasks
get_todays_tasks      → today's active tasks
create_task           → create a task with full metadata
update_task           → update any field
complete_task         → mark done (auto-spawns next occurrence for recurring tasks)
snooze_task           → snooze until a date
search_tasks          → full-text search
queue_agent_job       → dispatch a Claude Code agent to work on a task
list_habits           → all habits with today's completion status
log_habit             → mark a habit done or skipped
end_of_day_triage     → review and plan tomorrow
stale_backlog_review  → surface tasks that haven't been touched recently
... and more

A built-in terminal with docked and fullscreen modes — run Claude Code alongside your task view without switching windows.

Agents — any folder with an agent.config file becomes a dispatchable agent. Assign one to a task, queue a job, and Qalatra runs it with the task description as the prompt. Agents can write output files that preview directly in the app. Agents can also define output_rules to extract data from their output and automatically attach links back to the task.

Habits — recurring behaviors with flexible scheduling. Daily, weekdays, specific days of the week (Mon/Wed/Fri, Tue/Thu). Streak dots, session notes, completion history.

Recurrence — full RRULE support plus shorthands. Completing a recurring task auto-spawns the next occurrence.

Contexts — organize tasks by area of life or work. Each context has a color. Claude uses contexts when triaging and creating tasks.

Attachments — connect any S3-compatible bucket (Cloudflare R2 recommended) for file storage.

Theming — light, dark, or system. Full color token customization.

---

Upgrading from Task OS

If you were running the app when it was called Task OS, your database is stored under the old app name and needs to be moved once before launching Qalatra for the first time. The database itself is unchanged — this is just a folder rename.

Mac

cp -r ~/Library/Application\ Support/Task\ OS/db \
      ~/Library/Application\ Support/Qalatra/

Windows (run in PowerShell or Command Prompt)

xcopy "%APPDATA%\Task OS\db" "%APPDATA%\Qalatra\db\" /E /I

Then launch Qalatra. Your tasks, notes, habits, and history will all be there.

If you skip this step Qalatra starts with an empty database. You can still do it after the fact — just quit the app, run the copy command above, and relaunch.

---

Install

Download the latest release from the Releases page:

Platform	File
Mac (Apple Silicon)	.dmg (arm64)
Mac (Intel)	.dmg (x64)
Windows	.exe installer
Linux	.AppImage

Windows note: The installer is unsigned — Windows SmartScreen will warn you. Click "More info" → "Run anyway". If Claude Code fails on Windows, install it via npm: npm install -g @anthropic-ai/claude-code

---

Run from source

Requires Node.js 20+.

git clone https://github.com/pirateandfox/qalatra.git
cd qalatra
npm install
npm install --prefix ui
npm run electron-dev

---

Connect Claude

Qalatra runs an MCP server on http://localhost:3457. Add it to ~/.claude.json:

{
  "mcpServers": {
    "qalatra": {
      "type": "http",
      "url": "http://localhost:3457/mcp"
    }
  }
}

Restart Claude Code. The app must be running for the tools to be available.

You can also change the port and auto-update ~/.claude.json from Settings → MCP Server inside the app.

---

First-time setup

Open Settings (gear icon, top right) and configure two things before you start:

1. Working directory

Set this to your main projects folder — the directory where your work lives:

/Users/you/IdeaProjects        # Mac/Linux
C:\Users\you\IdeaProjects      # Windows

This is the working directory for the built-in terminal and the root Qalatra scans for agents. Everything flows from here.

2. Agent command

Qalatra dispatches agents using a configurable CLI command. The default is:

claude --dangerously-skip-permissions

Qalatra works with any command-line agent, not just Claude. If you use a different AI CLI, a custom script, or your own agent runner — change this to whatever command invokes it. Claude is what we use and what all the examples show, but it's not hardwired.

Per-agent overrides are also available — each agent folder can specify its own command in agent.config, so you can mix different agents across different tasks.

---

Agents

An agent is a folder with an agent.config file:

{
  "name": "Research Agent",
  "description": "Researches a topic and produces a markdown report",
  "command": "claude --dangerously-skip-permissions"
}

Qalatra scans your configured agents root and lists discovered agents in Settings. Assign an agent to a task from the detail panel. When you queue a job, Qalatra spawns the agent in that folder with the task description as the prompt. Results appear as a note on the task.

Agents can attach output files to tasks via the update_task MCP tool:

update_task(task_id: "...", links: [{ url: "/absolute/path/to/report.md" }])

Linked .md files open in a markdown editor with PDF export. Linked .html and .eml files open in an email preview.

Output rules — agents can parse their own stdout and automatically add links to the Qalatra task. Add an output_rules array to agent.config:

{
  "name": "My Agent",
  "command": "my-cli --flag '{description}'",
  "output_rules": [
    {
      "pattern": "Task ID: ([a-f0-9-]{36})",
      "action": "add_link",
      "url": "https://yourapp.com/tasks/{1}"
    }
  ]
}

pattern is a JavaScript regex. {1}, {2}, etc. reference capture groups. When the agent job completes successfully, Qalatra applies each rule against the raw output and adds matching links to the task. Rules live in your agent's repo — nothing is configured globally in Qalatra.

Starter agents: Clone taskos-projects-template as your working directory to get a set of ready-to-use agents — research, writing, triage, and more. Point Settings → Working Directory at the cloned folder to activate them.

---

Task fields

Field	Description
Title	What needs doing
Status	active, snoozed, done, archived
Context	Area of life/work (personal, work, project-x)
Project	Optional grouping within a context
Due date	Overdue tasks appear in red
Priority	p1–p4 for ordering within a section
Energy	high, medium, low — useful for matching tasks to your current state
Type	task, event, reminder
Recurrence	RRULE string or shorthand (daily, weekdays, weekly, monthly)
Agent	Path to an agent folder — enables job dispatch from the detail panel
Notes	Thread of notes with user/AI attribution

---

Keyboard shortcuts

Key	Action
N	New task
R	Refresh
Ctrl+`	Toggle terminal
Escape	Close panel

---

Tech stack

• Electron + Vite + React + TypeScript
• SQLite via better-sqlite3 (local, no server required)
• MCP via @modelcontextprotocol/sdk (StreamableHTTP, port 3457)
• xterm.js for the terminal
• S3-compatible storage for attachments (optional)

---

Contributing

git checkout -b feature/my-feature
# make changes
git commit -m "feat: what you did"
git push -u origin feature/my-feature
gh pr create --base main

Releases are cut by pushing a version tag: git tag v1.0.x && git push origin v1.0.x. This triggers the GitHub Actions build — macOS DMG (signed + notarized), Windows EXE, Linux AppImage — and publishes to GitHub Releases. The app auto-updates.

See CLAUDE.md for architecture, schema, and development notes.

---

License

MIT