Tools

A tool is something an agent can do rather than say — search the web, write a file, open a pull request, send a Slack message. An agent's tools decide the shape of what it can be asked for.

Who can do what

Task	User	Agent owner	Admin
Use the tools an agent already has	✅ shared in	✅	✅
Add or remove an agent's tools	— as Editor	✅	✅
Set a tool's parameters and credentials on an agent	— as Editor	✅	✅
Connect the organization-wide services tools depend on	—	—	✅
Register custom tools	—	—	✅

Tools are configured per agent, in Agent → Settings → Tools. Some need a connection an admin has set up first — an agent can't use the GitHub tool until GitHub is connected for the organization.

Always available

Every agent gets a core set with nothing to enable — the ones that make an agent an agent rather than a chat window:

Web	web_search, web_fetch, web_crawl, web_download — find, read, crawl, and download from the internet
Files	read_file, write_file, edit_file, list_files, search_files, view_file, send_file — work in the agent's file area, including sending a finished file to you in chat
Documents	create_document (Markdown, text, HTML), pdf_download (fetch a PDF and read its text)
Shell	run — execute a command in the agent's own sandbox
Memory	manage_memory, search_all_memories — save and recall memories, including across the organization's other agents
History	search_sessions — search its own past conversations
Scheduling	manage_routines — create and manage its own routines
Product docs	search_documentation, read_documentation — read these pages, which is how an agent answers "how does Delegate do X?"
Self-management	update_agent_settings, update_task_state, get_user_context — adjust its own name and greeting, keep a scratchpad through a long task, and know what time it is where you are
People	email, request_credential, invite_user — ask you a question by email, request a secret through a secure form, or invite a colleague
Agents	call_agent — hand work to another agent in the organization and wait for its answer

Connection-backed

These need an organization connection before they work:

Tool	Needs	Does
github	GitHub connection	Browse repos, clone, commit, push, open pull requests
slack	Slack connection	Send messages and files, list channels and users
gmail_send, gmail_read	Google connection	Send and search mail
google_workspace_admin	Workspace Admin connection	List, search, create, and update users in your Google Workspace domain
docworker	DocWorker connection	Build document knowledge bases and search them semantically
manage_identity	—	Read and update the user and company identity
manage_skills	—	Create and manage skills
manage_agents	—	Create and configure other agents
ssh	Your own credentials	Work on a remote server over persistent sessions

Configuring a tool

Each tool has its own Parameters & Credentials on the agent. Some are configuration (a default repository, a channel); some are secrets.

Secrets are encrypted at rest, and an agent never reveals its own credentials — if you ask it to print an API key it holds, it can't. An agent can also ask you for a credential mid-task via a secure input form, which is the right way to hand one over: it goes straight into encrypted configuration rather than sitting in the conversation.

Bundles

Related tools are usually enabled as a group. See Tool bundles.

Your own tools

If the capability you need isn't here, expose it yourself — an HTTP endpoint or an MCP server your agents can call. See Custom tools, and Exposing your tools over MCP for how to build the server.

Fewer tools is often better

Every tool an agent holds is described to the model on every task. A long list costs tokens and adds ways to go wrong. Give an agent the tools its job needs and leave the rest off; make a second agent for a second job.