Task Composer
The Task Composer is where you define a new task. A well-constructed task gets better results — this page explains every option available.
Overview
A new task combines an outcome, the files it needs and choices about how the agent should work. Choose the Project first so the task has the right files and context, then describe the output you want.
| Composer choice | What it controls |
|---|---|
| Prompt and attachments | The goal and source material |
| Agent preset | The model and its defaults |
| Workflow | Standard execution, research, review or Agent Swarm |
| Tools and skills | The actions the agent can use |
| Cloud or Local, in the desktop app | The computer that executes commands; it does not move workspace files |
Use the defaults for a first task. The sections below explain individual choices when you need to change them.
To send later, click the arrow beside Send, choose Schedule send, then pick a date and time and confirm. You can also right-click Send to open the menu. The time uses your device's timezone. The task waits until that time before starting, even if you close the tab.
Writing an effective prompt
Your prompt is the agent's instruction set. Structure it like this:
- State the desired output first. For example: "Generate a CSV summary of monthly sales by region."
- Add constraints. Format requirements, file names, style guides, deadlines.
- Provide context. If the agent needs to know about your stack, business logic, or prior decisions, include a brief description or attach a reference file.
- Be specific about scope. Ambiguity leads to the agent making assumptions. Spell out what's in and what's out.
Example of a vague prompt:
"Analyze the data."
Example of a strong prompt:
"Read
data/sales_2025.csv, calculate total revenue per region for Q1, and write the result tooutput/q1_summary.csv. Use Python with pandas. Skip rows whereregionis blank."
Selecting an agent
Below the prompt field, you can choose an agent preset. Agents control the reasoning effort and model used for the task:
| Agent | Best for |
|---|---|
| Default | Most tasks — balanced speed and quality. |
| Deep Think | Complex problems requiring deeper reasoning (slower, higher effort). |
| Fast | Quick tasks where speed matters more than thoroughness. |
If you're unsure, leave it on Default. See Agents & Personalities for more detail.
Some presets may be visible only to super-admin accounts, depending on how your platform is configured.
Choosing where the task runs
In Meowbert Desktop, a server-hosted workspace defaults to Cloud. Once your server operator has enabled local execution and your computer has access to the workspace's shared storage, choose Local · your computer's name in the composer. Then choose Sandbox for Docker Desktop or Host for execution directly on your computer. Local tasks use the Standard workflow.
Switching to Local does not download or move the server workspace. A workspace created On this device always targets its owning computer and has no Cloud/Local selector. See Local tasks in the desktop app for setup, storage requirements and mode permissions. The browser version does not offer the Local selector.
Choosing a workflow
Open the task parameters menu to choose the task workflow before sending it.
Standard task
This is the default for most work. Use it when one agent run is enough.
Long Horizon
Choose Long Horizon when you want the agent to:
- clarify first,
- write a durable plan,
- execute against that plan,
- and pass a reviewer panel before the task is allowed to finish.
This is ideal for larger tasks where you do not want the model to stop after a shallow partial answer.
Deep Research
Choose Deep Research when you need a careful, source-aware answer. It follows the Long Horizon clarify, plan, execute, and review flow, but is designed to investigate both deeply and broadly as the question requires.
For example, it can follow official documentation and PDFs closely for a specific program, or compare official information with independent reporting and user experience when the real-world picture may differ. It asks concise questions when the research goal or expected deliverable is unclear, keeps meaningful progress updates during the work, and by default gives a concise conclusion first with the sources it actually used at the end.
Turn on Web Search in Tools when you want current web research. The task can also use any other tools, Skills, or Sources you select.
Quality control (Experimental)
Choose Quality control when the final wording, structure, formatting, or visual design needs dedicated review before delivery.
Quality control follows the same clarify, plan, execute, and review flow as Long Horizon. Its reviewer focuses on the human experience of the result: whether the reply is intuitive and well organized, whether text artifacts read naturally, and whether presentations, documents, images, and other visual files are clean and polished.
If the reviewer finds a problem, the task revises the result and submits it again before finishing.
Agent Swarm
Choose Agent Swarm when the leader should be able to split the work across copies of itself, get several independent answers to a judgment call, and have its reply checked before delivery. Picking a swarm in the model list sets this up for you.
The form starts from the swarm's own settings:
- Lane model router, shown when your server has one, lets the leader run some lanes on other models.
- Effort controls how readily the leader delegates: lower values wait longer and keep the fan-out shallower; higher values nudge earlier and allow a wider, deeper split within the Advanced limits. Balanced keeps the normal defaults.
- Budgets, off by default, give the leader and its lanes one shared token budget and an optional time budget.
- Advanced... holds the lane limits, such as how many lanes run at once and whether replies are checked automatically.
This works well when:
- the work has parts that can be done at the same time,
- you want independent answers compared instead of one chain of thought,
- or you want a fresh reviewer to read the reply before you get it.
See Long Horizon, Deep Research, Quality control & Agent Swarm for the full breakdown.
Incognito chat
Use Incognito chat when you want a short-lived task that should not appear in project task lists or task search.
Incognito chats behave like normal standard tasks while they are open, but they are intentionally temporary:
- They do not appear in the project task list.
- They do not appear in workspace or project task search.
- They are automatically deleted after 24 hours.
- You can delete them permanently from the Task Detail toolbar.
Incognito chat is available for standard tasks. It is not used for scheduled tasks, Long Horizon, Deep Research, Quality control, Agent Swarm, or tasks attached to a project canvas.
Attaching files
Click the attachment icon to upload files into the task. These are placed in the task's working directory so the agent can reference them directly. Use this for:
- Source data files (CSV, JSON, logs).
- Reference documents the agent should read.
- Templates or existing files to be modified.
If your workspace has a connected file Source such as Google Drive, OneDrive, pCloud, or rclone, the same attach menu also lets you bring files in directly from that source.
The attach menu can also list project Canvases. Attach a canvas when the task should reference an existing interactive website, or choose Canvas mode when you want the next task to continue editing that canvas.
For Google Drive, OneDrive, pCloud, and rclone, the picker can offer two modes:
- Copy — import a one-off local copy into the task's
inputs/folder. - Live sync — create a local working copy in
inputs/, but keep it linked to the same remote file or folder so you can pull newer remote changes or push your finished edits back later from the Files browser.
The Live sync option appears only when the workspace's source connection has write access. In Live sync mode, folders can be linked too; pushing a linked folder can create new remote files/folders and remove remote children that were deleted locally. When a live-sync item is attached, Meowbert keeps the agent aware that it is working on a linked source path rather than a detached copy. For native Google Docs, Sheets, and Slides, Meowbert uses an exported Office file as the local working copy and asks for confirmation before creating the live-sync link. pCloud and rclone search match file and folder names/paths.
Attaching Google Docs, Sheets, and Slides
When a Google Drive selection includes a native Google Doc, Sheet, or Slides presentation, Meowbert asks how to attach it:
- Convert to Office files — Recommended and selected by default. Creates DOCX, XLSX, or PPTX files that use the existing Copy or Live sync workflow.
- Attach Google files directly — Experimental. Adds a reference to the original Google file and lets the agent read and edit that shared file immediately.
Direct attachments modify the original Google file rather than a local working copy. They are available only when the connected Google Drive account has write access, and they do not use Live sync Pull or Push controls because changes are applied directly.
Choosing direct attachment adds that Google file to the Project's Google Workspace allowlist. The file is then available to Google Workspace tools in every task in that Project, including when it was originally attached from a task rather than Project Context. Connecting Google Drive alone does not authorize every file in the account, and copying a .gdoc, .gsheet, or .gslides pointer into another Project does not grant access there. For a mixed selection, this choice applies only to native Google Workspace files; other files keep the selected Copy or Live sync mode.
If your admins have enabled YouTube, the attach menu also lets you paste a video link. In that case, Meowbert attaches a compact metadata summary for the video instead of downloading the media itself.
File size limits apply per upload. For large datasets, consider uploading to the project first (via Files browser) and referencing the path in your prompt.
Mentioning files and skills with @
Type @ in the prompt to point at something directly. A menu opens above the cursor listing the task's attached files and the available skills, each labelled Attached file or Skill. Keep typing to filter the list, use the arrow keys and Enter to pick an item, or press Esc to close the menu and keep the @ as plain text.
The menu opens when @ starts the prompt or follows a space, so email addresses don't trigger it. Your pick appears in the prompt as a pill, and what it does depends on its type:
| You mention | What happens |
|---|---|
| An attached file | The agent sees exactly which file you mean at that point in the sentence. Clicking the pill in a sent message opens the file. |
| A skill | The skill turns on in Tools, and the agent is told you called it directly. |
For example: "Use @Humanize to rewrite the introduction of @draft.docx."
A skill you mention stays on for later messages until you turn it off in Tools. It can't be turned off while its pill is still in the prompt, so delete the pill first. To remove a pill, select it and press Backspace.
Other ways to cite a file
Attaching a file makes it available to the agent, while citing it also marks where it belongs in your prompt. You can cite a file without typing @:
- Right-click an attached file above the prompt and choose Cite to insert a pill at the cursor.
- Paste a file into the prompt to upload it and insert a pill at the cursor.
Files you drag onto the composer are attached without a pill. Removing an attached file also removes its pills from the prompt.
Tool toggles
Click the wrench icon to expand the tool options panel. These toggles enable additional capabilities for the task:
| Toggle | What it enables |
|---|---|
| Web Search | The agent can fetch current information from the web during the task. |
| Memory | The agent can semantically search the workspace's persistent .memory directory during the task. This toggle appears only when Memory is enabled for the workspace. |
| Schedule Tasks | Enables recurring/scheduled execution. A schedule picker appears in the prompt area. |
| Subtasks | The agent can delegate scoped sub-tasks to child tasks, useful for large multi-step workflows. |
| Skills | Enables workspace-defined Skills (MCP tool packs). A skill selector appears to choose which skill to activate. |
Sources are selected from the separate Sources button next to Tools.
When you enable a source there, the agent can use source-specific tools during the run — for example, searching Google Drive, pCloud, or rclone, fetching a Drive, OneDrive, pCloud, or rclone file into the task filesystem, searching Outlook emails, viewing calendars, safely managing Outlook events, or using yt-dlp workflows for YouTube links and metadata.
Outlook is a tool-only source rather than a file-import source. That means it does not appear in the Attach files picker, but it does let the agent work against the connected mailbox and calendars during the run.
If the task also includes live-synced source files, the agent gets dedicated sync actions for those linked task files so it can check status, pull the latest remote version, or push finished local edits back safely.
When a workspace has Memory enabled, new task drafts start with Memory turned on by default, but you can disable it for any specific run.
These tool selections also carry into workflow runs. For example, if you start an Agent Swarm with Web Search and Memory enabled, the leader and its lanes receive those same selected capabilities.
Workspace owners connect sources from Connectors → Sources. Server admins can configure the provider settings by following Sources (Admin Setup).
See Tool Options Reference for details on each option and Skills for the available skill packs.
Canvas
Turn on Canvas in Tools when you want Meowbert to show you pages and visuals, inline in its reply or beside the chat. To keep working on a project canvas, attach it from the attachment menu; those tasks do not run in Quick mode.
See Canvas for how canvases appear and how to keep one in the project.
Submitting the task
Click Send (or press Ctrl+Enter / Cmd+Enter) to queue the task. You'll be taken to the Task Detail view where you can watch the agent work in real time via the events stream.
From there, you can also open sidebar threads on assistant replies when you want to ask a focused follow-up without interrupting the main conversation. See Task Follow-up for the full thread workflow.
If you prefer a different message shortcut, open Preferences from the sidebar. You can switch between the default Enter to send behavior and Shift+Enter to send.