taskuary

Taskuary

English · 简体中文

Your work manager. Locally. Securely. In control.

A personal AI assistant for your job. Mail, chats and tickets become tasks, the agents you already use (Claude Code, Codex, Gemini) do the work, and nothing goes out until you approve. Open source, runs on your machine, no subscription.

Taskuary is early—currently v0.3.9.34—so breaking changes are still possible before 1.0.

What Taskuary can do

One request, from arrival to your approval. Follow Ruth's request for the latest vendor spend numbers through the real app, using fictional demo data.

1. Everything lands in one place

Mail, chats, issue trackers, alerts, and reports arrive on one rail, sorted by what they need: Tasks (every open task - a spinning ring on the one an agent is working, a quiet shade on an urgent one), For later, Reports, the Advisor's ideas, and FYI. Beside it, the day opens on one card: what changed since last night, what was learned, the morning digest, your meetings, and what waits in each section - so you never open each system in turn. A goal is a task whose steps are tasks: triage opens one for an open-ended ask, and files later mail under the step it serves.

2. Let the Assistant walk you through it

Choose Walk me through my tasks. The Assistant puts one item on the table at a time, as the same task card you would open yourself, with the next actions under it: Next, Write reply, Send to agent, Mark done. The buttons always do the same thing; the chat line is for everything else, and the Assistant already has the message, the thread, and your past decisions to answer from.

Assistant look-ups clamp result limits to a safe range and use their default when a limit is invalid.

3. Every request becomes a task with its whole story

Open any item to see how it got here: the message, what triage decided and why, which agent has it, and where the reply stands.

4. The agent works inside the chat

Send to agent and it works in the same card. Here the analyst prepares the numbers, checks that the categories add up, names its source, and drafts the reply. Nothing is sent.

5. The last word is yours

The prepared reply waits on the task, beside the request that started it. Read it, edit it, and choose Approve & send when it is ready.

6. Start your day before it gets busy

Each morning the Assistant opens on what changed since last night, what it learned, today's meetings, and what waits in each section - and the morning digest unfolds in place, so it is never stale by the time you reach it. Ruth's request has a clear deadline: the 11:30 operations review.

Key features

Use your coding CLI

Connect Claude Code, Codex, Qwen Code, OpenCode, Kimi Code, Gemini, Cursor, Copilot, Muse Code, or another CLI. Set up the connection once, then give your agents profiles with their own instructions. Follow their sessions, answer questions, and review the result from Taskuary.

For Qwen Code, see the setup and compatibility guide. Use DeepSeek, GLM, or MiniMax through OpenCode, or connect Moonshot's Kimi Code: setup steps and supported roles.

A shared Hub for what agents learn

Keep discoveries, decisions, and useful warnings by topic. Agents can find what earlier work uncovered, discuss it, and correct it instead of starting from scratch.

Agents leave notes for each other

The Board's Live handoffs show what agents are working on, what is blocked, and what is ready. An agent leaves a note; the next one reads it before picking up the work.

Memory that learns how you work

Every draft you edit and task you reclassify gives Taskuary evidence about your preferences. Repeated patterns become lessons in LEARNED.md: how you write, what you own, and what deserves a task. For example, repeatedly moving the numbers to the top of a reply can teach it to lead with the total next time.

Open Docs → LEARNED.md to read, edit, or delete those lessons. Your written instructions in SOUL.md take precedence.

It also learns from what came after. The outcome ledger records one row per action Taskuary took or proposed - a reply sent, a task closed, an agent stopping and your next move in your own words, a pull request merged, a promise made, a sender ignored - and a daily sweep fills in what followed from facts already in the store: more rounds on the thread, a task reopened, a promise delivered late. Reflection reads those rows as evidence, so a lesson can say "green PR from the coding agent: you merge it, 8 of 8". A learned line never acts on its own; it changes which button comes first. It also keeps a page per person it deals with, compiled from the same ledger: how fast they answer, whether a chase is needed, who on copy moves them - and a thread gone quiet past that person's own day is slipped, on your rail under Tasks, with Chase, Let it go, They answered elsewhere and Remind me under it. See The Advisor.

• Learn from a correction. A model call turns an explicit correction into a hypothesis. Batched reflection compares multiple decisions; untouched approvals contribute here.
• Keep the evidence. Each machine-written lesson carries s (strength), ev (evidence IDs), and seen (last supporting date). A stable k identifies it across rewrites.
• Promote supported patterns. A new hypothesis starts at strength 2. Reflection is instructed to promote it at 4 or more, with at least three episodes across two people or threads. Contradictions weaken it; stale hypotheses also decay across reflection cycles.
• Use active lessons. The prompt builder excludes the Hypotheses, Proposed rules, and raw Verdicts sections. Active lessons inform triage, drafts, and agent context.
• Keep the owner in charge. Inferred rules that hide or file work wait in Proposed rules. Two matching explicit owner verdicts can already supply that authorization. Untagged lines you write are preserved, and learning can be disabled in Settings.

See Learning from your decisions, or the implementation in learn.py and learnedgraph.py.

What leaves your machine

Taskuary runs locally. With a hosted AI provider or coding CLI, the prompt contains the context needed to do the work you asked for. Use a model running on your machine, and those AI prompts stay local too.

Credentials are taken out of that prompt first. If a colleague mails an API key, a connection string, or a private key, it is replaced with a labelled placeholder ([redacted:aws-key]) at each of the three doors a prompt can leave by: the hosted models, a headless CLI run, and the first prompt of an agent pane. Your mail itself is never altered—the scrub is on the way out, not on the way in—so a vendor's one-time code stays readable where it arrived. Nothing Taskuary sends carries a placeholder either: a reply still holding one is refused, not delivered.

The rules are deterministic rather than a model's judgement, because by the time a model could judge, the credential would already be in a prompt. So they catch credentials with a recognizable shape—provider keys, tokens, credentialed URLs, connection strings, private