What Revolve Agent is

A coding agent that lives in your terminal: you type a message, a language model reads and edits the files in your project, runs commands, and answers.

You start it in the directory you want worked on. Every message you send starts a turn. During a turn the model can read files, search them, write them, run a command, fetch a web page, or hand part of the job to a second agent, and it keeps going until it has an answer for you. Then it waits for your next message.

Three things separate it from a plain chat window with a file picker.

The files it may touch are decided by the process, not by the model

Every path the model asks for is resolved against the directory the program was started in. A path that leaves that directory is refused, and so is anything inside .git. Commands do not go through a shell: there are no pipes, no &&, no $(), no redirects — the model names a program and a list of arguments, and a rule file decides whether that program may run. See Permissions for what that does and does not protect you from.

A session is a file, and it outlives the window

Conversations are written to disk message by message. You can close the terminal and resume where you were, name a session and come back to it tomorrow, or detach a session so it keeps running with nobody watching and attach a terminal to it later. See Sessions, memory, goals and Scripting and embedding.

The terminal is one of five front ends

The same agent runs behind a full-screen terminal interface, a line mode for piped input, a one-shot JSON mode for scripts, a JSON-lines protocol for editors and wrappers, and a socket for detached sessions. Nothing you can do interactively is missing from the scriptable modes, including answering a permission prompt.

Where to go next

If you want toRead
Get it running and send a first messageInstall and first run
Know every flag and subcommandCommand line
Know what to type into the chatChat commands
Know what the model itself can doWhat it can do
Control which commands it may runPermissions
Teach it your project's conventionsSkills and packages
Let it work while you are awayWorking unattended
Drive it from a programScripting and embedding
Point it at a model providerConfiguration

A word on what it is not

Revolve does not host models. It talks to a provider you configure — an Anthropic-compatible or an OpenAI-compatible HTTP endpoint — with a key you supply. Without a configured provider or an API key in the environment it has nothing to ask.

The permission engine is not a security boundary either. It catches the stray rm -rf and the command meant for another repository. It does not contain code that the agent itself wrote and then compiled; Permissions explains why and what an actual boundary would look like.