Taskuary Work Assistant
Turn work messages into tasks, hand them to the coding agents you already use, and approve every reply before it goes out.
- Source repo
- ldbumble/taskuary
- Stars
- ★ 138
- Last updated
- 1d ago
- License
- MIT
- Primary language
- Python
- FA score
- 79/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Individual contributors who need a daily view across Outlook mail, chats, and tickets can sort incoming work in one priority-based timeline.
- Not a fit if
- Teams that require all AI prompts to stay on the machine
- Organizations that need stable interfaces before 1.0
- Source review
- 79/100 · Good
What does this agent do, and when should you use it?
Taskuary is a local personal AI assistant that brings mail, chats, tickets, alerts, and reports into one work timeline and sorts them by priority. Its web interface and optional desktop window show each task’s original message, triage decision, assigned agent, and reply status. You can hand work to coding CLIs such as Claude Code, Codex, and Gemini, follow their sessions, answer questions, and review their results from the task card. The Assistant can prepare a reply, but you must review and approve it before it is sent. It runs through Python on Windows, macOS, and Linux, with a single-file Windows app and Docker deployment also documented; the project is pre-1.0 and may still make breaking changes.
On first run, connect an AI provider or local Ollama model, at least one inbound channel, and a coding CLI to receive tasks. Taskuary brings mail, chats, tickets, alerts, and reports into a timeline, sorting items into groups such as Urgent, On you, For later, Advisor ideas, and FYI. Open a task to see its message, triage reasoning, agent, and reply status; choose Send to agent to have the agent work in the same card, where you can follow the session and inspect its output. Taskuary can learn preferences from your edits and reclassifications, recording repeated patterns in LEARNED.md. You still read and edit the final reply and choose Approve & send.
- Individual contributors who need a daily view across Outlook mail, chats, and tickets can sort incoming work in one priority-based timeline.
- Users who want Claude Code or Codex to handle email requests can follow the agent session and review its result in the task card.
- Teams that want agents to reuse discoveries and decisions can save and discuss topic-based knowledge in the Hub.
- Users who need agents to pass work and context to one another can leave progress, blockers, and handoff notes in the Board’s Live handoffs.
- People who want control over outgoing AI drafts can review and edit replies on the task before approving them for sending.
How do you install or deploy this agent?
Python 3.10 or newer is required; the documented Python install works on Windows, macOS, and Linux:
pip install taskuary
taskuaryThe app opens at http://127.0.0.1:7787. For an optional desktop window:
pip install "taskuary[desktop]"
taskuary-desktopFor Docker deployment:
git clone https://github.com/ldbumble/taskuary
cd taskuary
docker compose upThen open http://127.0.0.1:7787. Windows users can also download and open the single-file Taskuary.exe without installing Python.
How do you use this agent?
After starting the app, connect an AI provider or local Ollama model, add at least one inbound channel, and choose a coding CLI to receive tasks. The setup wizards test each connection before enabling it. Supported coding CLIs include Claude Code, Codex, Gemini, Qwen Code, OpenCode, Kimi Code, Cursor, and Copilot; you can give agents their own instructions. Open a timeline item and choose Send to agent to follow the work and inspect its result in the task card. To send a reply, review and edit the draft, then choose Approve & send.
What are this agent's strengths and limitations?
- Combines mail, chats, tickets, alerts, and reports in a single timeline sorted by priority.
- Supports a range of coding CLIs, including Claude Code, Codex, Gemini, Qwen Code, and OpenCode.
- Keeps agent work beside the original request and requires user approval before sending a reply.
- Offers a local-model path and scrubs recognizable credential patterns from prompts before they leave for a model or CLI.
- When using a hosted AI provider or coding CLI, the prompt context needed for the requested work is sent to that service; AI prompts stay local only with a model running on your machine.
- The README does not state specific model or hosted-provider prices; users may incur separate service costs.
- First-run setup requires a model, at least one inbound channel, and a coding CLI, so it involves more steps than a single install command.
- The project is at v0.3.7.17 and may make breaking changes before 1.0.
- With Docker, coding CLIs and the optional WhatsApp bridge still run on the host.
How does this agent compare with similar options?
The README compares Taskuary with the hosted assistant Fyxer: Taskuary is free, open source, runs on your machine, and hands work to agents you already use. It does not provide a feature-by-feature comparison or a detailed comparison of data handling.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Taskuary Work Assistant This agent | 79 · Good | CLIFree + model costs | ★ 138 | 1d ago | Python | Codex · Claude Code |
| AgentCall join-meeting Skill | 46 · Major gaps | Agent plugin / skillFreemium | ★ 165 | 27d ago | Python | Codex · Claude Code |
| Alook AI Workforce Layer | 42 · Major gaps | CLIFree + model costs | ★ 1.4k | today | TypeScript | Codex · Claude Code |
| Vellum Assistant | 56 · Major gaps | Desktop appFreemium | ★ 1.4k | today | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence shows a localhost binding by default, token and host restrictions, deterministic outbound credential scrubbing, audit logging, and approval before a draft is sent. The security policy also discloses plaintext local storage and that a CLI may act with the user's permissions. Least privilege, data-flow disclosure, and confirmation are handled well. Deductions reflect reliance on account and disk protections for stored secrets, pattern-based scrubbing that cannot detect natural-language secrets, no dependency vulnerability review in the supplied material, broad script/SQL execution by connectors and reports, and limited recovery guidance. Repository attribution is visible, but publisher identity is unverified.
README, packaging configuration, and CI descriptions are broadly consistent; CI covers multiple platforms, builds, browser checks, and Docker scenarios, and dependency ranges and some optional integrations are documented. Deductions reflect the early release and broad lower-bound dependency ranges, with no evidence here establishing dependency availability or actual run results. Failure guidance appears in the security and install notes, but user-facing handling for the range of runtime errors is not demonstrated in these files.
The product clearly targets personal work and covers mail, chat, tickets, coding CLIs, reports, and local models, with Windows, Python, Docker, and desktop entry points. The security policy describes its local single-user assumption, network exposure risks, and agent permission boundaries. Deductions reflect broad dispatch, connector, and CLI capabilities whose trigger conditions and individual boundaries are not fully specified in the supplied material.
The README is organized around features, installation, and documentation; commands, the local address, first-run setup, and demo limits are concrete. The license is complete, and the security policy states supported versions and limitations. Deductions reflect only partial evidence for naming stability, FAQ coverage, and change history; the project is pre-1.0. Maintainer and security-reporting paths are described, but publisher identity and more specific ownership responsibilities are not verified.
The evidence presents a centralized task workflow with preserved context, agent collaboration, draft review, and user-approved sending. The outputs are inspectable and editable, and the workflow has clear practical value. Deductions reflect no quantified time savings, operating costs, or integration maintenance costs in the supplied material; the breadth of features may add setup and upkeep burden.
Major product and security claims point to concrete documentation, configuration, and implementation paths; the README links learning claims to implementation, while packaging and CI provide installation, build, and check details. The security policy distinguishes facts such as plaintext local credentials, default unauthenticated access, and CLI permissions from their risks. Deductions reflect that much of the supporting evidence is repository-authored, with no independent corroboration in the supplied files; badges alone do not establish verified results.
- Credentials are stored in plaintext in local database and configuration files; protect them with the machine account and disk encryption, and avoid syncing the data directory to an untrusted location.
- Do not expose the service to an untrusted network. The API, reports, tools, and terminal can execute actions, and auto-approved agents may edit files, run commands, or push code.
- Scrubbing targets recognizable formats and cannot be relied on to catch natural-language passwords or every secret; review agent context and drafts before sending.