Your first task

This page takes you from connecting GitHub to a draft pull request. To try Cloud Agents without a repository, start from scratch instead; it needs no GitHub connection.

Before you start

  • Agent access. Cloud Agents are enabled per account during the beta.

  • A GitHub repository you can push to. The result is a branch and a pull request in that repository.

  • Daily quota. The agent’s model calls count toward your account’s daily usage quota, like your other API traffic.

1. Connect GitHub

  1. Open https://agents.freeinference.org (also linked as Agents from the FreeInference dashboard) and sign in with your FreeInference account.

  2. Choose Connect GitHub, either from Getting started in the sidebar or from My Settings → Integrations, and approve the sign-in on GitHub.

  3. Install the FreeInference GitHub App on your account or an organization, and select which repositories it may access.

Back on the Agents page, the repository picker lists the repositories the App can access and you can push to. To add more later, choose Configure repository access at the bottom of the picker, or Configure repositories on the Integrations page, then Refresh the picker.

To start in a new repository, choose Create new repository in the picker. It is created in your personal GitHub account, private unless you choose Public, with a README so it has a branch to start from. If the App does not have access to it yet, give it access and refresh.

To disconnect, choose Manage → Disconnect GitHub on the Integrations page. That removes your stored GitHub authorization. Uninstalling the App in your GitHub settings is what removes its access to your repositories.

2. Describe the task

In the composer:

  1. Describe the change the way you would brief a colleague: what to change, where, and how to tell it is done. For example: “Fix the SSE total-timeout regression on /v1/messages and add a unit test that reproduces it.”

  2. Pick the Repository and the Branch to start from.

  3. Pick the Agent, the Model and, where offered, the Reasoning level — see Agents and models.

  4. Choose Run, or press Enter. Shift+Enter adds a new line.

Cloud agents do best on self-contained changes: a bug fix with a clear description, a focused refactor, tests for an under-covered module, a documentation sweep. Put conventions the agent should always follow, such as how to run the tests, in your repository’s AGENTS.md or CLAUDE.md, or save them as memory.

Keep these limits in mind:

  • One repository per task, and GitHub only. GitLab is not supported yet.

  • Dependencies install over HTTPS. pip install, npm ci and Git over HTTPS work. The agent runs as an unprivileged user, so apt install does not; use a virtualenv or pip --user. Other ports, including Git over SSH, are blocked.

  • Each turn has a time limit of one hour.

The + menu in the composer attaches files, sets a goal, and picks the skills, MCP servers, SSH servers, plugins and environment for the task — see Working in a task.

3. Watch it work

The first turn takes longer: the platform creates the conversation’s sandbox and clones the repository into it. The agent keeps going if you close the page, and a blue dot appears beside the task when it finishes.

The task page streams the agent’s work as it happens. Chat shows its replies with compact tool calls; Trajectory shows every step with its input and output. If the agent asks you a question, a dialog opens and the turn waits for your answer. Under each finished turn, Recorded changes lists the files it changed.

4. Create the pull request

When you are happy with the changes, choose Create PR. The platform checks the changes, pushes them to an agent/… branch in your repository and opens a draft pull request, linked from the task. Your CI runs on that branch as on any other, if your workflows are set up that way.

Want changes? Reply in the same conversation. When the new turn finishes, Update PR pushes its changes to the same pull request. Merging stays with you.

Next steps

  • Working in a task — follow-ups, queued messages, stopping, goals, forks, and reviewing the pull request

  • Environments — install dependencies and start services before every task

  • Sandbox and security — what the agent can reach and what gets published