Equipping Tools

Tools are an agent’s capabilities. Too many slows it down and raises cost; too few prevents it from finishing the job. Picking the right set matters.

How to enable tools

1. Open the agent → Settings tab
2. In the Enabled Tools section, toggle the ones you need
3. Use the Auto Execute / Approval Required button next to each tool to set whether the agent must request approval before running it (risky tools default to Approval Required)
4. Save

Next to the section title, the number of currently enabled tools is shown like 32/45. Each tool row has an on/off checkbox and — for enabled tools only — an Auto Execute or Approval Required button.

The same Settings tab also has an External MCP section further down. Instead of toggling individual tools, you connect per server — pick an external MCP server you’ve already registered under Settings → MCP Servers and connect it to this agent. A connected server’s tools don’t appear in (or count toward) the Enabled Tools list above (e.g. 32/45); they’re managed separately, and each server has its own Auto-approve tool calls switch — leave it off and every call from that server’s tools needs approval.

💡 Recommended: 30 or fewer tools per agent. If you need more, break the work into a team — each AI employee can carry a different tool set.

Common tool categories

Category	Tools
🌐 Info gathering	Web Search · HTTP Request · Web Browser
✍️ Output	Ask User · Share File · Canvas Create/Update · Canvas Read · Canvas List
💾 Memory/files	Read File · Write File · Edit File
💻 Code/system	Bash · Computer Use
⏰ Automation	Schedule Create/List/Update/Delete · Task Create/Update/List · Update Plan
👥 Team collaboration	Team Agent List · Team Chat Send/Read · Team State Get/Set/List/Delete · Sub Agent Spawn
🧩 Skills	Skill Search · Skill Install/Activate · GitHub Skill Install · Skill Load · Skill Export · Skill Feedback · Skill Patch
🎨 Creative	Image Generate · Agent Set Avatar
🔐 Connections & credentials	Save/Delete Credential · Authorize Webapp Access · Authorize External API Access

How detailed are list/search results?

Tools that scan through several items at once behave differently depending on how often their results get acted on right away. Schedule List and Team Chat Read normally show just the key facts, then fetch the fuller version — with whatever detail a follow-up needs — once the agent decides it’s about to act on one of those results (updating a schedule it just looked up). Canvas List, Task List, and Skill Search default to the fuller version already, since acting on a result immediately (installing a skill it just found) is the common case for these. Either way, the agent makes this call mostly on its own, but you can also nudge it — asking for “just the summary” or “the full details” gets you the corresponding version.

🌐 Information gathering

Settings label	When to use
Web Search	When you need to search the web by keyword (with optional domain allow/block lists). Automatically combines multiple search engines and breaks complex queries into sub-searches, returning results with citations
HTTP Request	When you know the URL and need to call an external API or fetch a static page — supports custom methods, auth headers, and request bodies
Web Browser	Sites that require login, clicks, scrolling, or input. Whether a click auto-proceeds or shows an approval card depends on what the target element does — a “Post” or “Pay” button always asks; opening a menu does not. A logged-in site’s session can be saved straight into secure storage for reuse, so the password or login values never appear in the conversation or logs. See Handling Approvals for the full breakdown

Tip: For quickly pulling information, web search is the lighter, faster option. Reach for Web Browser when you need to work a site directly — login, clicks, or typing. You can also grant both and let the agent pick whichever fits the task.

✍️ Output

Settings label	When to use
Ask User	Ask the user a question mid-flight and wait for the answer. When it needs a sensitive value like a password or API key, the question card also shows an Answer with secure input button, so you can hand it over without ever typing it into the chat box
Share File	Share a generated file as a download link. Regular attachments like images stay free and permanent up to 10MB; documents (PDF/PPTX/DOCX) up to 10MB; other archives and data files (zip/tar/parquet, etc.) up to 200MB; video up to 1GB. Video is always billed in SU by size and auto-deleted after 14 days regardless of size; other files are billed and expire the same way only once they pass 10MB. Either way the original is removed after 2 downloads
Canvas Create/Update	Show markdown/code/document content — or a full webpage or slide deck — in the in-app canvas panel (same title → new version)
Canvas Read	Pull the current content of a canvas the agent already created — used before a partial edit, to see the full text again
Canvas List	List the canvas documents the agent has created
Send Card	Always-on plumbing — not shown in settings. Attaches a structured card (info card, OAuth connect, file picker, etc.) alongside a message, available across every trigger type

Canvases now come in just two forms — document and slide deck (the old standalone “webpage” type has been folded into documents). A plain document or webpage with no slide markers is automatically treated as a page — one continuous, scrollable view.

Tip: For a page-style canvas, use the export button in the top-right corner to save it as HTML, PDF, or PNG — and in the web app you can view and edit the source directly, or double-click text in the preview to edit it inline (mobile apps are view-only). A slide-deck canvas works differently: hit Edit and a dedicated slide editor opens, with your slide list on the left and a working canvas on the right. Draw and place text boxes, images, and shapes freely; adjust color, font (20 options, including 11 Korean and 5 Latin web fonts), alignment, and formatting; drop in tables and charts; apply layout and theme presets; undo/redo and copy/paste; and reorder slides by dragging their thumbnails — it handles like a real slide-design tool. Each slide can carry its own presenter notes, and once you’re done you can play the deck right there or export it as PPTX, PDF, or PNG. Vertical formats like Instagram-story ratios are automatically laid out correctly too (mobile apps are view-only for the slide editor as well). Either way, when you edit a canvas, the agent picks up your changes the next time it works on it. A single canvas can hold up to roughly 8 million characters; saving is rejected past that, so trim images or split the document if you hit the limit.

A canvas card in a chat bubble isn’t a live link that reloads the newest content every time you click it — it’s a snapshot of what existed the moment that card was created. When an agent revises the same document several times, cards keep piling up in the conversation, and clicking one opens the version it was created from, not today’s latest, so you can see exactly what a given pass changed. While you’re viewing a past version, the edit and delete buttons disappear and only a Back to latest button remains; you can still tap the version indicator at the top to open the version history and jump to other points in time. Pick an older version and hit Restore this version to bring it back as the current one. Version history doesn’t grow forever, though — everything saved in the last 14 days is kept, and anything older than that is kept only for the most recent 20 versions (a version you edited yourself, and whichever version is current right now, are always kept regardless of age or count). Opening a version that’s aged out shows the current version instead, with a note explaining why.

If the same canvas gets edited and saved from another device or another chat session first, trying to save from the screen you’re on shows a “this canvas was updated elsewhere” notice. You get to choose: Reload (replace your screen with what was just saved), Overwrite with mine (save your edits as-is), or cancel (dismiss the notice and keep editing) — nothing is silently dropped either way. Web, iOS, and Android all work the same way here.

💾 Memory & files

Settings label	When to use
Read File / Write File / Edit File	Work with the agent’s working files and persistent storage
Memory Scan	Always-on plumbing — not shown in settings. Lets the agent list its own accumulated memories
History Search	Always-on plumbing — not shown in settings. Searches set-aside chunks of earlier conversation from long sessions by keyword; the body of a match is then pulled back with Read File above
Source Chat Read	Shared snapshot sessions only — not shown in settings. Reads the original user’s conversation up to the moment the snapshot was taken, useful when a visitor asks about decisions or facts that predate the shared session
Snapshot Transcripts Read	Dynamically provided to agents that have published at least one shared snapshot. Review the conversations external users had through your shared links to improve future versions

Tip: Recall happens automatically on every turn. For agents that repeat work, keeping Read File / Write File enabled helps them build on what they’ve learned.

💻 Code & system

Settings label	When to use
Bash	Run a one-shot shell command in an isolated sandbox (data wrangling, file processing, simple API calls). Stdout and stderr are returned separately
Computer Use	Control a GUI desktop environment and shell. Includes terminals and general GUI apps, not just browsers

⏰ Automation

Settings label	When to use
Schedule Create / List / Update / Delete	Let the agent register and manage its own runs. Supports recurring schedules (e.g. every day at 9 AM) and one-time delayed runs (e.g. run once in 30 minutes)
Task Create / Update / List	Explicitly track multi-step work with dependencies and completion state — for commitments that span multiple sessions
Update Plan	Automatically maintains a step-by-step checklist for whatever the agent is working on right now. Remembers finished steps so it doesn’t re-fetch the same source or re-read the same file. Where Task (above) is a lasting, cross-session commitment, this is a lightweight checklist scoped to the current piece of work only
Open / Dispatch to / Complete Workstream	Always-on plumbing — not shown in settings. Keeps recurring work that needs to be revisited in a separate workstream instead of mixing it into the main conversation. See Working Across Workstreams for details

👥 Team collaboration

Settings label	When to use
Team Agent List	Discover other agents in the team (for @mentions)
Team Chat Send / Read	Post to the team chat. @AgentName wakes up that agent automatically. When a mention requires a reply, the mentioned agent sends a return mention once it finishes. A thread reply (a reply to a specific mention) can wake the original sender even when the mention link only points the other way
Team State Get / Set / List / Delete	Shared memory the whole team can read and write (separate from each agent’s private memory). Only the last agent to write a value, or the manager, can delete it
Sub Agent Spawn	Spin up a one-off helper agent and collect the result. Only available to employees who aren’t on a team

Tip: There’s no tool for adding an AI employee — press + Hire on screen and someone new starts right away, asking you about the job themselves. There’s no “equip” step for an existing employee either — every employee has every tool from day one. See Setting Up a Team for the full flow.

🧩 Skills

Settings label	When to use
Skill Search	Search Explore in the Skills tab or the agent’s own Library
Skill Install/Activate (⚠️ approval required)	Add a skill found via Explore to this agent (risky tool — Approval Required by default). Registered as a reference to the original without copying content; if you have previously customized the skill, your version is used
GitHub Skill Install (⚠️ approval required)	Install a skill directly from a public GitHub repo (owner/repo) you point at. The first call previews SKILL.md and the file tree; calling again with the same repo actually installs it. Like Skill Install/Activate, it carries higher risk, so no auto-approve scope is ever offered — every call needs its own one-off approval
Skill Load	Deploy an installed skill’s files into the sandbox and load its instructions. The agent then runs it itself with Bash
Skill Export	Automatically packages sandbox code into a reusable skill and saves it as a private draft. A review card arrives right away — nothing becomes public until you publish it from that card
Skill Feedback	Reports a defect in a skill the agent just used; for someone else’s skill it goes to the author as an improvement suggestion, and for a skill you have customized yourself it points you to fix it directly
Skill Patch	Self-edits a skill you have customized via exact-string replacement; saves a version before and after the change so you can roll back to the earlier one
Skill Inspect	Always-on plumbing — not shown in settings. Runs automatically the first time Skill Install is called, previewing a skill’s full description file (SKILL.md) and bundled file list before installing

🎨 Creative

Settings label	When to use
Image Generate	Text→image, reference-image edits (up to 14 reference images), multi-turn iterative refinement. Four resolution tiers (512·1K·2K·4K) and 14 aspect ratios available. A generated image keeps working as a real file in later turns — the agent can process it with Bash, embed it in HTML, or share it as a download, and it can drop the image straight into a Web Browser upload field or a file dialog opened through Computer Use (e.g. “put the logo you just made into this upload form”)
Agent Set Avatar	Change the agent’s profile image via AI generation or URL. Any agent can update its own avatar

🔐 Connections & credentials

The screens for storing and managing credentials are covered in full in Connecting External Services. Here’s just what each tool does.

Settings label	When to use
Save / Delete Credential	Store a value the agent received in chat (or obtained itself) under a name for long-term reuse, or remove it. Once saved, the agent can reuse it even during scheduled/autonomous runs when no one is watching
Authorize Webapp Access (⚠️ approval required)	Asks the owner to let a webapp the agent built and published call an external service using a stored account/credential. The approval card summarizes exactly what it would be allowed to do
Authorize External API Access (⚠️ approval required)	Asks the owner to let the code-execution (Bash) tool call an external service directly using a stored account/credential — approved the same way, with a summary card

Approval (human-in-the-loop)

Each tool has its own Auto Execute / Approval Required setting. Skill Install/Activate and GitHub Skill Install have their per-tool approval setting turned on to Approval Required from the start, so the agent has to wait for you to approve from chat before running it.

Skill Export is not approval-required — it runs automatically. The moment the agent calls it, a private draft skill is created and a review card arrives with it, and nothing becomes public until you publish it from that card.

Beyond that, a few more tools ask for approval by default because of what they do. Share File, Sub Agent Spawn, Schedule Create/Update/Delete, Authorize Webapp Access, and Authorize External API Access raise an approval card before running by default — the last two carry even higher risk, so no auto-approve scope is offered for them at all: every call needs its own one-off approval. HTTP Request (requests that send or change data), Computer Use (actions that change the screen), and Web Browser (actions with lasting effects like posting, paying, or deleting) also ask depending on how risky the action is. You can still tune these through their per-tool setting or your auto-approve settings. See Handling Approvals for the full breakdown.

Recommended combinations

📰 Information curation agent

Web Search is enough — autonomous runs deliver their results straight to chat automatically regardless of what triggered them, plus a push notification if you’re not looking at the app.

📊 Monitoring agent

HTTP Request + Memory Scan + Write File + Schedule Create/List — track changes over time.

📝 Report writing agent

Web Search + Write File + Canvas Create/Update — research, write, and display in-app.

🔧 Data wrangling agent

Bash + Write File + Share File — process data and deliver as a download link.

Too many tools?

The more tools you enable, the longer the LLM spends reading every tool description per turn — slower responses and higher cost. Aim for 30 or fewer per agent.

If your task is so complex it needs more, split into a team. Each AI employee can have a different tool set, and they coordinate through team chat.

---

Advanced

Read this only if you want to know exactly how approval decisions are made, or why some tools never show up in Settings.

• How approval is decided: Each tool judges its own risk from the call itself. Web Browser, for example, only asks for approval when the element it’s about to click does something hard to undo — “Post,” “Pay,” “Delete” — and lets ordinary clicks, like opening a menu, proceed automatically. On top of that, your per-tool Auto Execute / Approval Required setting applies, along with whatever auto-approve scope you picked from an approval card (next N calls, for the rest of this task, or always). Bash loses only the “always” option, for safety — the other scopes stay available. Skill Install, GitHub Skill Install, Authorize Webapp Access, and Authorize External API Access carry even higher risk: no scope picker appears for them at all, so every call needs its own one-off approval.
• Always-on built-in tools: Send Card, Memory Scan, History Search, and the skill-preview step that runs automatically the first time you install a skill are wiring every agent carries by default. They don’t appear in Settings and can’t be turned off.
• Tools only visitors get: In a shared snapshot session, a couple of extra tools attach automatically for the external visitor — reading the original conversation up to the share point, and (once you’ve published at least one shared link) reading back what visitors discussed through your links. Neither shows up in your own Settings list.
• Retired tool names: The old Notify User, Team Agent Create, Team Agent Equip, and History Read tools no longer exist. Notify User was replaced by automatic delivery — autonomous runs now reach you in chat and, via push if you’re not looking at the app, regardless of what triggered them. Creating an AI employee moved to the Hire button on screen (the manager’s own creation tools, Create Agent and Create Team, were retired alongside it on 2026-08-21); the separate tool for adding tools to an existing employee was dropped entirely once every employee got every tool from day one; and History Read was absorbed into Read File. Older conversation history may still show these retired names.

Next steps

• Installing Skills & Presets — Use packages where the tool combination is already tuned
• Skills & the Autonomous Skill System — Save proven workflows as reusable skills. Covers how Skill Feedback and Skill Patch connect to the autonomous skill system in detail.