Abilities

An ability is a set of tools chloe ships. Each is true or false in abilities in the agent's agent.ts, with nothing to import: abilities: { memoryPerUser: true, scripts: true }. Every tool chloe ships has one, so anything built in can be switched on from the dashboard, which writes the same line, says only what differs from the default, and puts the file back with the reason when the agent would not load. A tool you write yourself, and a connection's, is bound in tools instead (Tools).

The tool names here are the ones the model sees. Your purpose and skills should say what to do ("note it in your memory"), not name these tools, so a rename breaks nothing (A prompt with tools).

The abilities

Ability	Default	Tools it adds
selfImprovement	on	selfWriteFile
memory	on	memoryListFiles, memoryReadFile, memorySearchFiles, memoryEditFile
scripts	off	scriptRun
memoryPerUser	off	memoryWriteUserNotes
folders	off	folderListFiles, folderReadFile, folderSearchFiles
web	off	webReadPage, webSearch
googleMail	off	emailReadMail, emailReplyMail, emailSendMail
googleCalendar	off	calendarListEvents, calendarAddEvent
googleDrive	off	driveSearchFiles, driveReadFile
resend	off	resendSendEmail

memory

The agent's notes between runs, in its memory folder and nowhere else (An agent says where that is). false leaves them out.

Tool	What it does
memoryListFiles	Lists the files in its memory.
memoryReadFile	Reads one.
memorySearchFiles	Finds the files that contain some text. The search is plain text, not by meaning.
memoryEditFile	Makes one or more changes to a file in one call, all or none. Each replaces one piece of text, which must be there once (the spaces at the start and end of its lines may differ). With no old text it adds to the end of the file, or makes the file. There is no tool that replaces a whole file, because that is how the rest of one gets lost. It also moves, renames or deletes a file or a folder, after any changes and in the same commit.

Every run that changes the memory is one git commit under the agent's id.

folders

Folders the agent may read and never change, like a project's code or the docs of a product it answers questions about. They are yours, not its memory. paths is the full path of each, and the model knows each by the last part of its path, so no two may end the same; about is a few words for them.

abilities: { folders: { paths: ["/home/you/shop", "/home/you/shop-site"], about: "the shop's code and its site" } }

Tool	What it does
folderListFiles	Lists a folder in one of them. The model is shown each one's layout, two levels down, at the start of each reply.
folderReadFile	Reads one file, or some of its lines.
folderSearchFiles	Finds some text in one of them, with the lines around each match.

In a git repository, what git ignores is left out, so .env.local, dist/ and anything else in .gitignore is never listed, searched or read. A new file git would add is seen at once. Any .env, .git, secrets and node_modules is left out everywhere. A folder that is not there, or a path that is not full, stops the agent loading. Somebody you invited gets these tools when you give them chat; a guest on a web channel gets them only when the channel names them.

web

webReadPage reads one public web page as plain text, with links written as [text](url). It refuses this machine and private addresses, and does not run a page's JavaScript. webSearch searches the web and gives back results, each a title, a url and a line, which the model reads with webReadPage: it goes to Brave Search when connections.brave.api_key is set, and to DuckDuckGo when it is not. web is true for both; switch webSearch off on its own to read pages only.

resend

resendSendEmail sends one email through Resend to the addresses in to, from from. The model writes only the subject and the body. when says when it should send, in your words; replyTo, tag, markdown and keep are as in Tools. Needs the Resend connection: a key in connections.resend.api_key. Without one nothing is sent and nothing fails.

googleMail, googleCalendar and googleDrive

Google, as three abilities that share the one sign-in (see Connections). Each is true for its narrowest reach, or its fields written out, which the Abilities page shows as a form. The model never chooses what a field fixes. They are named for Google because that is all chloe ships: another mail, calendar or drive is a tool you bind yourself, or a mailbox of your own on the email channel.

Ability	Fields	Tools
googleMail	search (a Gmail search, default in:inbox), days (7), about, sendTo (none), sendFrom, sendWhen	emailReadMail, emailReplyMail; with a sendTo, emailSendMail, which writes only the subject and the body
googleCalendar	calendars (["primary"]), days (7), about	calendarListEvents, calendarAddEvent, which invites nobody
googleDrive	search (all of Drive), about	driveSearchFiles, driveReadFile

abilities: { googleMail: { search: "label:orders", days: 3 } }. What an ability cannot hold (a copy folder for sent mail, several calendars to add to, a limit worked out in code) is bound with the tool functions in tools, which are on Tools. An email channel that may write first brings emailStartConversation itself: see start in Email, and a Telegram channel that may write first brings telegramSendMessage: see send in Telegram.

selfImprovement

selfWriteFile changes files in the agent's own folder: its purpose, skills and jobs, and its code (agent.ts, tools, services, channels, scripts). It writes every file one change needs in one call, and that is one commit you can read and undo from the dashboard. With code among them, the agent is loaded with the new files first, and if it would not load, every file is put back and the agent is told why.

It is on by default, which suits an agent that is still being shaped. Switch it off once the agent does what you want. It never writes its evals or its memory. Code it writes runs with the same rights as chloe, so its folder bounds what it writes, not what that code does.

It is in a turn only when the agent's owner asked, and is refused once another tool in that reply read from outside the agent. The whole rule is What an agent may change about itself.

memoryPerUser

memoryWriteUserNotes keeps one note on each person the agent talks to, on any channel, as users/<channel>-<id>.md in its memory. The note is shown at the top of each turn with that person, and the tool replaces it. Whose note it is comes from who sent the message, never from the model, so one person on two channels has two notes. A visitor on a web channel gets this tool even though the channel does not name it.

scripts

scriptRun runs any file in the agent's scripts/ folder, or any executable file inside one of its skill folders (named <skill>/<path>), and returns what it printed, or lists them. Each script must be executable (chmod +x and a #! first line; one the agent writes itself with a #! is made so), runs inside scripts/ or its skill's folder, and gets the memory folder in MEMORY_FOLDER. With this on and no script in either, the agent does not load. Give each script a skill that says when to run it, or keep it in the skill's folder. With it on, a file somebody sends is also saved for the scripts (Skills and scripts).

Switching one tool off

An ability is all of its tools. To keep an ability but not one of its tools, or to switch off a tool you bound yourself without losing the options you gave it, list its name in without: without: ["memoryEditFile"]. The dashboard writes that list when you switch a single tool off: on the Abilities page for one an ability adds, and on the Tools page for one you declared. A name that is no tool of the agent stops it loading, saying what there is. An ability's tool may be named while the ability is off.

On the Abilities page the group and its tools agree: with every one of an ability's tools switched off, the ability reads off too, and switching it on brings all of them back (its tools are taken out of without).

With no ability

Tool	Who gets it	What it does
chatSendFile	a conversation on a channel that can send files, never a guest's	Sends a file from its memory, or one it was sent, into that conversation. Telegram can.
ownListFiles	every turn	Lists the files in its own folder, and which it may change. Never its memory.
ownReadFile	every turn	Reads any of them, several in one call: its own source code and its skills included. The agent is shown every skill's description and the file to read, and reads the one it needs.
ownListRuns	turns its owner wrote	Lists its own runs, newest first: the job or chat, where it came from, when, the cost, and whether it failed.
ownReadRun	turns its owner wrote	Reads one run: what started it, the answer or the error, and each step on the way.
ownTryJob	turns its owner wrote	Runs one of its jobs as a trial, as its files are now: it reads for real, sends nothing, and answers with every message and email it would have sent.
ownReadGuide	turns its owner wrote	Reads these guides, several in one call, for the version installed, so it can say what it could be given (Gmail, a channel, web pages) and how. Its files and the list of guides are in its purpose in those turns, and an agent that may change itself is told to read what a change needs at once and never to ask for a secret in a chat.

The owner is owner in settings, the first entry in a channel's allowFrom, or the account on the dashboard. Reading its own folder is on for every turn, but a job, a token and a visitor never get the tools that read its runs or try its jobs, because a run holds what other people said. Reading a run, or trying a job, counts as reading from outside, so a change after it in the same conversation is refused.

On a channel that names its tools

A channel's tools keeps a turn to the tools it names, plus the memory tools, ownListFiles and ownReadFile (an agent may always read its own folder). A visitor on a web channel is a guest and gets only what the channel names, and memoryWriteUserNotes with memoryPerUser on. Naming a memory or self tool there stops the agent from loading. See Channels.