Tools
What a bot can do on its Runner, including files, search, the shell, the web, teamwork, and what it brings to you.
A bot works on its Runner with the same kind of tools a coding agent gets. It runs as you, with your user's permissions, on the computer you assigned it to.
On the computer
| Tool | What the bot does with it |
|---|---|
| Read, write, edit | Opens, creates, and changes files. |
| Grep, find, ls | Finds the line, the file, or the folder before touching anything. These respect .gitignore. |
| Shell | Runs commands: builds, tests, git, scripts. Stop kills the running command. |
Relative paths resolve in the bot's working directory.
Commands see the same PATH and variables as your terminal. On a Mac or Linux, Lorca reads them from your login shell, whether zsh, bash, or fish, when it starts, so npm, cargo, and brew work even when the app started Lorca. Plugin servers get them too. Restart Lorca after you change your shell's configuration.
A bot has your access
Shell commands run as your user and are not limited to the working directory. They can control files, processes, credentials, applications, and the network. Give a bot a description with constraints that match the trust you would give a script you run yourself, and assign bots to a machine you are comfortable with them controlling. A bot's Access can turn its shell commands off, or keep its files to reading.
Before a shell command starts, Lorca checks the bot's Access, then parses all of its stages and sends the complete command through Auto-review. Safe work runs automatically; for risky work, the command's card asks you first. An allowed command retains your user's full access.
The command's card
A command shows in the chat as a card only while it needs you: while Auto-review asks for your permission (Allow once, Always allow, or Deny), and once the bot leaves a running command to you, because its turn ended with the command still running or it waits on a question the command asks. That card shows the latest output and Stop. Until then, the shimmering line says what the bot is running. Click or tap the command to see all of it.
Running tasks
Every command the chat's bots have running shows under Running tasks, including the ones a bot is still dealing with. While a chat has one, a terminal button sits in the chat's toolbar, beside ⋯ on a phone; in the desktop app, a badge on it counts them when there are two or more. It opens the list: what each command does, who runs it, and for how long. The desktop app shows each command, its latest output, and Stop right in the list; on a phone, tap a command for them. Stop ends that command alone; the bot hears that it stopped and carries on. A command shows there once it has run for two seconds, so quick ones never do.
Servers and long builds
A bot starts a dev server, a file watcher, or a long build in the background. It reads the command's first two seconds of output, so a server that fails to start or says where it listens still reaches it, then goes on with its work, and the command keeps running after the bot's turn ends. It shows under Running tasks rather than in the chat, unless it asks a question. Stop in the desktop app's composer leaves it running; stop it from Running tasks. It isn't stopped for being quiet either. When it ends by itself, say a server crashed or a build finished, the bot hears how it went and carries on.
You can send one there yourself. While the bot is still waiting on a command, Run in Background sits beside Stop under Running tasks: the bot goes on with its work, and the command keeps running as if the bot had started it in the background. In the desktop app, Chat › Run Command in Background does the same for every such command in the chat, with ⌃B on a Mac.
When a command asks for input
Each command runs in a terminal of its own, so sudo, ssh, a key's passphrase, or a [Y/n] can ask for an answer. A command that stops at a question, or prints nothing for 20 seconds, doesn't hold up the bot: its turn goes on, and the bot answers what it knows. The rest is yours, on the command's card.
- Answer from the card, on any of your Devices. What you type goes straight to the command. It is never saved in the chat, and the terminal doesn't echo it, so the bot doesn't read it back. Bots are told to leave passwords and one-time codes to you. When the command finishes, the bot hears how it went and carries on.
- Stop from the card or from Running tasks. Stop in the desktop app's composer ends the whole turn, its commands included, except those running in the background.
Anything a bot types into a command goes through Auto-review first, like a new command. A command still waiting stops by itself after 30 minutes without output, when Lorca quits, and when its chat is deleted. On Windows, background commands and answers need Windows 10 version 1809 or later.
On the web
DeepSeek, Anthropic, ChatGPT, and Grok bots search the web and read pages when the answer is not on the disk. The search runs on the provider's side, so it needs no setup. Grok bots can also search X.
With the team
| Tool | What it does |
|---|---|
| Message a bot | Hands work to a bot outside the current chat, which reports back. See Chats. |
| List teammates | Sees every bot, its job, its Runner and model, and whether that Runner is online. |
| Create and edit bots | Builds the team: a new teammate, or a change to one's description, model, or working directory. |
| Tasks | Keeps work that takes more than one turn. See Tasks. |
| Routines | Schedules its own recurring work. See Routines. |
| Memory | Saves and recalls what it has learned. See Memory. |
| Skills | Reads its skills when a task matches, and drafts new ones for you to save. See Skills. |
| Project | In a group, reads the project and suggests entries. See Group chats. |
| Plugins | Searches the marketplace, asks to install a plugin, and asks you to sign in. See Plugins. |
| Coding agents | Passes a coding request on to Claude Code, Codex, OpenCode, pi, or fx, which reports back, and relays its report to you. See Coding agents. |
| Channels | Listens on Telegram or Slack when you ask. See Channels. |
With you
| Tool | What it does |
|---|---|
| Publish an output | Posts a file or a link it made, with what it checked. See Outputs. |
| Leave for review | Leaves an action or a draft for you to approve later. See Drafts and reviews. |
| Ask for a secret | Asks for a password or a key on a card, and uses it by name. See Secrets. |
| Attention | Reports what needs you, and posts a brief when it coordinates. See Attention. |
Plugin tools
Every plugin installed on a Runner makes its tools discoverable to the bots there, as far as each bot's Access allows: creating a GitHub issue, reading a Linear ticket, driving a browser. Lorca loads only the matching tool schemas when a task needs them. Actions that change something go through Auto-review first.