Code mode: run a coding agent on your own repository

A coding agent on your own repository #

Code mode gives you the Claude-Code-style workflow as a desktop app: sessions are grouped by repository, each session pins a working folder, a git branch and a model. The agent explores the repo with read tools — list, read, find, grep — before it ever proposes a change.

The desktop app is free. For a local Code run, model tokens are billed by the provider through your own API key. Cloud execution is the separate subscription layer; cloud features are temporarily free during the current beta, and More AI promises advance notice before billing begins.

From task to plan to build #

Turn on “Plan first” when you want a structured plan — summary, steps and touched files — to wait for “Approve & build” before edits. With it off, the selected access mode controls whether the agent proceeds and which actions require approval.

During the build you see every write as a diff card with added/removed line counts, expandable to the exact lines.

Commands, tests and the permission model #

The agent can run commands with output streaming into the transcript. Access modes range from read-only through auto to Full Access. Auto uses its configured command rules and approvals; Full Access turns off normal approvals and sandboxing, although catastrophic commands remain blocked and writes to protected files still ask.

Repo conventions are picked up automatically: if your project ships a CLAUDE.md, AGENTS.md, .cursorrules or similar instruction file, the agent reads and follows it.

Steering, sub-agents and questions #

You can type while the agent works — mid-run messages are injected into its next step, so course-corrections don’t require an abort. For wide research tasks the agent can fan out read-only sub-agents that explore in parallel and report back. And when information is genuinely missing, it asks you a structured question instead of guessing.

Branches, worktrees and pull requests #

Code mode supports isolated git worktrees, so the agent can work on a separate branch without changing your current checkout. It can also open a pull request through your local GitHub CLI when gh is installed and authenticated. You can review the branch, diff, commands and tests before publishing.

Choose the agent: More AI, Claude Code, Codex or OpenCode (coming soon) #

Code is getting a choice of coding agent. Alongside More AI’s built-in agent, a session will be able to run Claude Code, Codex or OpenCode. The feature is in testing and not part of the current desktop release; until it ships, Code runs More AI’s own agent.

More AI keeps everything around the agent: the session and its transcript, the access mode, the built-in browser, connectors, skills, memory and sync. The selected agent plans and writes the code; More AI checks each command against the access mode before it runs, and the agent’s own execution paths do not bypass those checks.

Each agent is installed and updated inside the app — not globally and not in your repository — and you can pin a version or roll back. It runs on a compatible API key you added to More AI or on the agent’s own sign-in: Claude Code needs an Anthropic-compatible API, Codex an OpenAI or OpenRouter key, and OpenCode works with most providers, including local models in Ollama and LM Studio.

You can switch agents inside a session: the new one continues the same conversation, but it does not inherit the previous agent’s internal context. An import arrives alongside, bringing skills, commands, instruction files and MCP connectors over from Claude Code, Codex, OpenCode, Cursor and Gemini; imported connectors stay off until you enable them. Claude Code, Codex and OpenCode are names of Anthropic, OpenAI and Anomaly Innovations; More AI is not affiliated with these companies.

Code: choose the computer or a cloud workspace #

Local Code works in a repository on the selected computer. Paired remote access, when available, sends instructions to that same computer and needs it online. Synchronizing its history does not grant remote execution permission.

Cloud Code is a separate beta workflow: use the account interface to check availability and connect a supported repository. Choosing Cloud creates a separate workspace; it does not move a running local task, upload an uncommitted checkout or copy local keys and tools. Review the files, tests and any publication action before sharing the result.

• Sync, cloud and other devices

Frequently asked questions #

How is Code mode priced?

The desktop app is free, and local Code runs bill model tokens through your provider key. Cloud execution belongs to the subscription layer; it is temporarily free during the current beta, with advance notice promised before billing begins.

Will it change my code without approval?

With Plan first enabled, substantial work waits for “Approve & build”. After that, the access mode controls approvals. Full Access does not pause for ordinary risky commands because normal approvals and sandboxing are off; catastrophic commands remain blocked and protected-file writes still ask.

Does it follow my repository’s conventions?

Yes. If your project ships a CLAUDE.md, AGENTS.md, .cursorrules or similar instruction file, the agent reads it and follows your conventions automatically.

Which languages and projects does Code mode work with?

Any repository on your disk. The agent works with your files and runs the commands your project already uses — installing dependencies, running your test suite — so it is not tied to a particular language or framework.

Can it open pull requests or work on a separate branch?

Yes. Isolated git worktrees keep a separate branch away from your current checkout, and More AI can open a pull request through your local GitHub CLI when gh is installed and authenticated. You still review the diff, commands and tests before publishing.

Does choosing Cloud move my local repository and running task?

No. Cloud Code uses a separate workspace and its own available repository access and model configuration. Local changes, processes, provider keys and desktop tools are not transferred automatically.

Can Code run Claude Code, Codex or OpenCode?

Soon. Support is in testing: a Code session will be able to use More AI’s built-in agent, Claude Code, Codex or OpenCode, with the same access modes, browser, connectors and sync. Until it ships in a desktop release, Code runs the built-in agent.