Everyday tools
The buttons in the top bar: Nova conversa (New chat) up front, search and, in one group, Command, Report, Teams and Board (on smaller screens, icon only; on the phone, at the bottom). All of them cover the machines and buildings you can see.
New chat
Nova conversa (New chat, or the n key) opens an empty tab with the message box already focused, like Claude Desktop: just type and send. The chat only exists once you send the first message; closing the tab before that leaves nothing behind.
The choices are chips around the box, already set to what you used last time:
- Project: starts at No project, a standalone chat to ask, write and think together, which lives in "Standalone chats", with history and search. Pick a recent project and the same screen opens a development session in its folder, with the Worktree option. Choose a folder… and Create a new folder… open the full form (folder browser,
git init), with what you already wrote. The browser shows only folder names, inside your home folder, and hides folders such as.ssh. - Agent: Claude, Codex, Antigravity or Copilot. With no project, picking one the machine doesn’t have switches to a machine that does; in a project, it’s whatever that project’s machine has.
- Model and what the agent can do (Ask for approval, Can edit files, Automatic, Plan only). The mode for standalone chats is kept apart from the one for project sessions.
- Machine (when there is more than one) and role (when you have roles).
Paste or drop images into the box, or dictate with the microphone. The session shows up in the office and the tab becomes the chat. A Claude session started from the dashboard also shows up in Claude Code’s conversation history in the project’s VS Code, so you can continue there.
In a new worktree: the session works in a copy of the repository, on a branch of its own (fleetly/<name>), next to the repository in <repo>.worktrees/. That way several tasks move forward in the same project without one agent touching another’s files. The root .env files come along; node_modules doesn’t. Later, in the Code tab, you open a PR, merge or discard (see Code and Git).
Roles
A role is a template for a seat on your team: the frontend specialist, the reviewer, triage. In Edit roles (Teams → Roles), or in the full new session form (Role → Create roles…; or Edit roles…), each role has:
- Role prompt: who it is, what it does and how it finishes. It is fixed and goes into the session’s instructions, separate from the task you type each time.
- Skills: what it should know or follow, each with a name, when to use it and the content. In Claude Code, Antigravity and Copilot, each skill becomes a
SKILL.mdfor that role only; in Codex, it goes along with the instructions. - Agent (Claude, Codex, Antigravity or Copilot) and model.
- Internet in Codex’s sandbox (Codex roles only): Codex driven by the dashboard runs without internet; when on, the role can install dependencies and start the app.
- Off limits: editing files, running commands or
git push. In Codex, which can’t block tools one by one, blocking edits makes the role run read-only, and commands and push go as instructions. In Antigravity and Copilot, Fleetly refuses each forbidden action. - Account connectors (Claude roles only): Choose the connectors lists the ones in your Claude account (Gmail, Drive, Notion…), each with three options side by side: Off, Read only or Read and write; at the top, how many are on. On Read only, the role reads without asking for approval and nothing is written there; on Read and write, anything that writes waits for you in the Outbox; the role doesn’t use the ones that are off. With no choice, the role sees them all and each use follows the usual approvals.
When you pick a role for a new session, it opens in the role’s agent and model, and the task you type is that time’s request. The role keeps applying to the next messages, and its name shows in the session header. There are ready-made examples to start from (frontend, backend, reviewer, triage).
Import Claude Code agents: in Roles, Look for .claude/agents on your machines lists the agents you already have (in ~/.claude/agents and in the .claude/agents folders of projects with recent sessions). Tapping one opens the role already filled in, with the prompt, the model and the tools left out as blocked, for you to review and save.
The role applies as it was when the session opened: changing the role applies to new sessions. Sessions opened in the terminal or VS Code don’t get a role after they start.
Teams
Agent teams have their own page: Agent teams.
Search
Shortcut /. Takes you to a session: type the project and the topic, in any order (for example, a1 deploy). First, right away, the sessions that match by name, summary, tags and what they did in the terminal; below, the ones that mentioned the topic in the chat (Claude and Codex, last 30 days), one line per session. With nothing typed, the ones you opened last. ↑ ↓ and Enter open.
Each session shows when it last did a deploy, push and commit (only commands that succeeded). Searching deploy brings the session that did the deploy first, not just the ones that talked about it; the project list shows the Last deploy at the top.
Today
In the Report, the Today tab shows what changed in the period, in time order: commits, edited files and each session’s status. Filter by project.
Summary
In the Report, Summary tab: Claude reads the sessions’ activity in the period (8 hours, 24 hours, 3 or 7 days) and sums up what needs you, what got done and what is still pending. It uses a little of your plan’s limit, on the machine it runs on.
Routines
Every routine in one place: in Teams, the Routines button (or the guide’s Schedule a routine step) lists the routines of every project and machine, with when they run, the next time and how the last one went, plus Run now. To create one, pick the project in New routine in and tap Create; to change one, Edit. Both happen right there, without leaving the list (← All routines goes back). The form has three parts: What it does, When and How.
Tasks that run on their own on a schedule, in a folder, always in a new session: for example, reviewing at night what changed during the day. They live on the project’s machine and run with the dashboard closed. Open the project list (the project name in the session header, or the room’s sign) and use New routine:
- What it does: Review what changed (the last 24 hours, without touching the code), Run the tests and fix, A feature from start to finish or Just my instructions. You can add extra instructions.
- A feature from start to finish: you describe the feature and the routine works in three steps, each in a new session in the same worktree: the agent builds it (plans, implements, tests and commits), another agent reviews it (Codex reviews Claude, and Claude reviews the others; with only one, a new session of it) and the first one fixes what the review pointed out. The routine’s row shows the step (building, reviewing, fixing).
- Who does it: Claude, Codex, Antigravity or GitHub Copilot. The agent must be installed on the folder’s machine; Fleetly tells you when saving if it isn’t.
- When: On a schedule (the days of the week and the time, on the machine’s clock; if it’s off at that time, the routine runs when it turns on, up to 2 hours later; with no day selected, it only runs with Run now or on leftovers), On every commit (when the chosen branch gets a new commit in the folder, at most every 10 minutes) or When a session finishes (you pick the session; the routine runs once, when it stops working, and then pauses).
- Open the pull request at the end: with a worktree, if the run ends with commits, Fleetly pushes the branch and opens the pull request with the machine’s GitHub CLI (
gh), and the routine shows Open the pull request. The agent still can’t push; you do the merge. - Run in a worktree: the routine works on a copy of the repository, on a branch just for that run, and can commit there; your folder stays untouched. If it finishes without changing anything, the worktree goes away on its own. Without a worktree, it works in the folder and doesn’t commit.
- Accept automatically: commands and edits go through without asking. Without it, each request waits for you up to 15 minutes (it reaches your phone) and is then denied.
- Account connectors (Claude routines): as in a role, choose which connectors from your Claude account the routine uses and how. The Read only ones are read without waiting for approval, so an overnight routine reads Gmail or Drive without stopping; nothing is written to them. The Read and write ones go through the Outbox.
- Time limit: the run stops on its own when it gets there.
Every routine has fixed rules: it never pushes or opens a pull request (the request is denied even with accept automatically) and the building policy still applies. Each run shows up in the routine, with Open the session to read what it did; the session gets the rotina tag. When the run finishes, your phone gets a single “Rotina terminou” (routine finished) notice, instead of one for each step. Creating, changing, running and deleting is for whoever operates the building and goes into the audit log.
Usage window leftovers: the plan’s 5-hour window resets and whatever is left is lost. Turn on Use leftovers on this machine (at the end of the Routines section) and mark routines with Can run on the usage window’s leftovers: close to the reset, with usage left over according to the forecast, the machine runs the one that hasn’t run for the longest, with a time limit that ends before the reset. You choose how much must be left and how many minutes before. It doesn’t run if the week is tight. Instead of Run on its own, choose Just suggest on my phone: you get a notice with the suggested routine, and tapping it opens Fleetly with Run (you can accept up to 10 minutes before the window resets). In the Report, the Usage put to work from leftovers line shows how many routines ran on leftovers in the period, with the time and tokens.
Report and budgets
The Report button brings together everything that happened, in tabs: Today, Summary, Projects, Weeks and People. In Projects, over the last 7, 14 or 30 days: time with agents working, tokens, cost at API prices, sessions, commits and edited files, with work per day. On Team and Enterprise, the People tab shows the numbers per person.
Weeks is the retro: the week so far and the weeks already closed, with sessions, time, commits and files, the projects that delivered the most, the routines that ran and the teams’ rounds and sends. It covers only the projects you see, and never by person. To get each week on your phone, turn on the weekly retro in Settings → Notifications and approvals.
Share (on each week) builds a card for LinkedIn or X: you choose what shows (commits, sessions, time, files, how many projects, the project names, routines, teams and a photo of your desk, if you set one up) and the post text comes with it. Project names start off. The card is built in your browser: nothing leaves until you Download image, Copy image, Copy text or open LinkedIn or X (the image is copied so you can paste it into the post).
With set budget, give a project a weekly limit in millions of tokens. The dashboard warns you when it passes 80% and 100%, on your phone too. On Enterprise, there’s also a monthly budget in dollars, with a lock (see Enterprise).
Open sessions
The button with the number of open sessions lists the sessions open on the machines (the running processes), with the memory each one uses and how long it has been idle. Select the idle ones and use End: it only ends the process, and the chat stays saved to resume from VS Code or the dashboard.
Automatic names and status
After each step, the cheapest model gives the session a short name and writes one line about where it stands. It shows up in every view. It uses a little of the limit; turn it off in Settings → Plan and sessions. It needs Claude Code on the machine.
Didn’t find what you were looking for? Write to admin@fleetly.tools.