Start from scratch

Start from scratch runs a task without a repository: a script, a prototype, or work on files you attach to your message. It needs no GitHub connection. Until GitHub is connected, or while the App can access no repository, the composer starts from scratch; otherwise choose Start from scratch in the repository picker. These conversations are grouped in a Start from scratch folder in the sidebar.

The agent starts in an empty workspace that belongs to the conversation. Follow-ups, Stop, changing the model, goals, forks, side chats, archiving and deleting work as in any other conversation, but nothing is published: there is no branch, no pull request and no Create PR. The first time you send a task from scratch, a notice says so; choose Got it to run it. It is not shown again in that browser.

A conversation from scratch has no repository environment. It starts from the platform’s base image with your personal global variables.

Keep your files

The files stay in the conversation’s cloud workspace, which is stopped when idle and started again when you continue. Permanently deleting the conversation, and every fork that shares its workspace, deletes the workspace and its files. Keep what you need first:

  • Download it. Download workspace on each turn’s result saves the whole workspace as one .tar.gz file. Git data (.git) is left out. Every other file is included, hidden files too, and .gitignore does not filter it. Files you attached are kept outside the workspace, so they are not in the download unless the agent copied them in. A workspace over the download limit is refused rather than saved in part. If that happens, ask the agent to remove what you do not need, such as installed dependencies or build output, and download again.

  • Put it on GitHub. Connect the GitHub MCP server, then ask the agent to create a repository and push the files to it. In the composer, open the + menu, choose MCP servers, then Add MCP, and pick GitHub. Choose Sign in with GitHub and approve access on GitHub; no token is needed. The GitHub server sends each file’s contents in a tool call, so this suits small projects of text files; download larger ones instead.

You can connect GitHub with a personal access token instead. It needs Contents read and write access to the repository. For the agent to create the repository as well, the token also needs access to all repositories and permission to create them; otherwise, create an empty repository on GitHub and give the token access to just that one.

If the agent finds no tool to push files, the GitHub connection is read-only: remove it and add GitHub again. Your GitHub access stays with the platform and does not enter the sandbox, but through it the agent can change any repository the connection can reach (see Sandbox and security). Pushing does not turn the conversation into one on that repository. To continue with pull requests, give the FreeInference GitHub App access to the new repository and start a new task on it.

Start in a new repository instead

If you already know the work belongs in a repository, choose Create new repository in the composer’s repository picker and start the task there. See Your first task.