Take the full Helios tour.

Full transcript

0:00 · Meet Helios

Helios is a native Linux workbench for coding agents. It puts conversations, project files, review tools, and local previews in one window. In this tour, we will walk through the main interface, then follow a small project through the tools you use around a coding conversation.

We are using a fictional task board called Sprout. The conversations, approval requests, and background tasks are staged examples. The application screens are real, and the local file, browser, terminal, and Git actions are real. This narration is AI generated.

Helios connects to different provider tools rather than making every provider behave identically. A control that exists for Claude may work differently for Codex, Muse, Kimi, Google, or OpenRouter. We will point out those differences.

0:48 · Install and connect a provider

Start on the Helios website and choose an installation path that matches your system. Ubuntu twenty six point zero four has a package route. The latest release has a Debian package; source instructions list the system dependencies. Check the version: an older stable package may not include everything shown here.

Open Settings, then Accounts. This clean demonstration home has no signed-in provider accounts, so the status rows explain what is missing. Claude Code, OpenAI Codex, Google Antigravity, Muse Code, and Kimi Code use their own installed tools and authentication. OpenRouter instead needs its API key in Helios's Providers settings.

You only need one provider to begin. Follow that provider's installation and sign-in instructions, return to Accounts, and refresh its status. In Providers you can inspect binary and model information. A subscription and an API key can have different billing rules, so use the account route you actually intend to use.

1:46 · A project, a prompt and attachments

Choose your project folder before the first message. A new installation can otherwise open in your home folder, which Helios treats as read only. The new-chat menu offers a project folder, and Settings has a default working directory. We will use the Sprout repository.

Write a concrete request in the composer. Control Enter sends it; Enter starts a new line. The plus button adds editable file-path references. You can also drop files or paste an image. Image payloads are supported for Claude, Codex, Muse, and Kimi; Google and OpenRouter currently keep the path as text.

The command button opens the available command surface. Native actions depend on the provider: Compact is wired for Claude, Codex, and Kimi, while native Review is a Codex action. Read those capability labels before relying on them. The response in this tour is staged.

2:38 · Find and manage conversations

The sidebar keeps saved conversations close to their projects. Select a row to read its transcript and restore that conversation's controls. Switching the visible chat does not automatically stop other managed sessions. Status indicators help you notice a conversation that is working or waiting for you.

Use the search button to search local session content. Here, searching for keyboard finds the conversation about shortcuts, including a matching snippet. Choose a result to jump back into that session. Use it when you remember a discussion rather than its title.

Right click a local session for Rename, copy actions, or Delete session. Renaming gives it a stable, useful title. Deletion is an explicit operation, with checks for sessions held elsewhere. Older inactive sessions can also move into Helios's archive. Archiving and deletion are separate.

3:30 · Branch a conversation and recover files

A conversation fork and a Git worktree solve different problems. Claude offers a fork from a message in the transcript. Codex and Kimi have a native fork from the current head through the command surface. They require a native provider session; here we show the entry points.

For separate files, choose the optional new-chat worktree action. Helios creates a Git worktree only when you request it. The new chat gets its own checkout and branch. The Changes pane offers removal for worktrees Helios owns, and refuses to remove one with uncommitted changes.

Rewind files opens checkpoints taken before turns in a Git project. Inspect the changed paths before choosing a restore. Checkpoints exclude ignored files and do not reverse external side effects. Stop an active writing response first; Helios blocks a rewind while that write is still in progress.

4:22 · Models, permissions and questions

The model menu sits directly above the composer. Its available rows depend on the installed provider tools and connected account catalogs. Open Execution to see the selected conversation's workflow, reasoning options, and permissions. Settings provides defaults, while conversation controls belong to the chat you are working in.

Permission labels reflect the provider's native contract. They do not mean every provider has the same sandbox. For example, Google's headless route has its own distinct modes and cannot answer interactive approval prompts here. Kimi's Never ask maps to its native unrestricted mode, so choose deliberately.

Here are a staged approval and question. Read the command or diff, the explanation, and the grant scope before responding. Question dialogs can offer choices or free text. Dismissing an approval does not grant consent. Real answers return to the provider that asked.

5:17 · Keep the objective and plan visible

For work that takes more than a quick question, open Set a goal. Give the conversation an objective, a clear Done when condition, and a short acceptance checklist. Here, success means rejecting empty task titles while keeping the existing behavior for valid titles. The goal belongs to you.

Save the goal to keep its status and checklist progress beside the composer. The controls let you pause, resume, edit, or mark the goal complete. Set that state from evidence you have reviewed, not just a confident final answer. Pause changes the goal; use Stop to interrupt a turn.

The Plan workspace pane shows the agent's execution plan and conversation-derived quality signals separately from your acceptance checklist. A completed plan step is useful progress information, but it is not independent proof that a test passed or a deployment succeeded. Local goals do not need tandem services.

6:11 · Context, capabilities and optional services

Open Claude memory in the workspace selector to inspect the selected project's context files. Helios shows the relevant Claude instructions and memory files, with an editor for existing files. These real files in the fictional home are Claude context, not universal provider memory.

Capabilities is the place to inspect the capabilities and configuration evidence Helios has for the current provider. Missing information should remain visible as missing. Use it to investigate unavailable models, tools, or native actions.

Two more workspace areas are optional. Shared connects to a configured network scratchpad, and Missions needs the external tandem command-line tool and its state. They are not configured here. A mission's Approve and continue action approves a gate; it does not automatically run the next leg.

7:01 · Review the actual Git changes

Switch to Changes to inspect the actual repository. The list includes staged, unstaged, and untracked files in this working folder, including changes from other sessions. It shows the worktree, not just changes made by the selected conversation.

Open a file to read its diff. You can stage or unstage a whole file or an individual hunk. The validation change is small: trim the title, reject an empty value, and leave the normal return value intact. Staging updates the real local Git index.

Review comments send specific feedback to the selected session, so check the comments before submitting them. Prepare a commit message after reviewing the staged result. Commit, Push, and Create PR are explicit repository actions. This demonstration repository has no remote, and we are not publishing from this tour.

7:52 · Preview and attach what rendered

The browser is a separate pane beside the chat. Open a project HTML file, a local development server, or another page, and keep the result in view while you discuss it. The browser has tabs, navigation controls, an address field, and its own profile separate from your regular browser.

Here we are opening a real local Sprout page. The camera button adds a screenshot of the visible page to the composer. You now have a concrete image to discuss. The composer's separate desktop screenshot button depends on a working desktop portal, which is unavailable in this headless environment.

Supported agent sessions can open, read, click, type, inspect console output, and screenshot through Helios's browser tools. Access outside the project's local scope asks for consent, and your own browsing is not automatically shared with an agent. Google's current route does not receive these per-session Helios browser tools.

8:48 · Use the terminal and run menu

The Terminal workspace pane gives you tabs of your own shell in the selected chat's folder. It starts a shell when you open the pane. This terminal is for you; the coding agent still runs commands through its own provider tools. Here we are running the project's real local test.

Select relevant terminal output and use Send to chat to place it in the message box. It does not send the message automatically. Use this to ask about a failure or share an exact diagnostic result.

The Run menu reads supported project launch configurations and package scripts. A configuration with a port can open its local preview after the server answers. Review commands and environment settings as project code. In this isolated demonstration, the menu and the local test show the workflow without installing or contacting a service.

9:37 · Background work, stopping and spend

Background shows tasks exposed by supported Claude and Muse sessions, such as shells, agents, workflows, or monitors. These rows are staged examples so you can see the controls. Output is available only when the driver provides a local output file. Stop asks that task's provider to stop; it is not a file rollback.

The application menu also has Emergency stop all sessions for sessions managed by this app. Check what actually stopped and review the worktree afterward. It does not cancel processes outside Helios, undo completed commands, or restore files.

Spend rules differ by route. Claude API processes have a ten-dollar cap, Codex API thread families have a two-hundred-thousand-token cap, and OpenRouter has a five-dollar allowance per Work with bounded tool rounds. Verified Claude and ChatGPT subscriptions disable the corresponding Helios API caps. Google, Muse, and Kimi have no equivalent Helios cap established here.

10:36 · Settings, diagnostics and your first project

Finish in Settings to set your default folder, model, and permissions. Appearance lets you follow the system theme or choose light or dark. The translucent sidebar is optional and depends on your desktop's appearance setup. Behavior includes notifications and the title-generation backend, whose default is First message, with no model call.

Other title backends have prerequisites: Ollama needs a running service, the built-in local engine needs its command-line runtime and model, and Claude uses its provider. We are only showing those choices. Providers and Tools help diagnose missing binaries, authentication, catalogs, and external tool connections; the log-file entry locates diagnostics, and About identifies the running version.

The easiest first project is a small one: choose a folder, connect one provider, describe one change, inspect its result, and run a check you understand. Add the other tools as you need them. The Helios website links the installation guide, provider details, source code, and a place to report problems.