Suite / Helios / Compared with a terminal
Helios vs. Claude Code in a terminal
Same CLI underneath. Nothing is taken away; some things stop being invisible.
If you already run Claude Code, OpenAI Codex or an OpenRouter model from a terminal, this page is the honest answer to one question: what actually changes if you open the same agent inside Helios, and when should you not bother.
The short version
Helios does not replace the claude or codex command. It launches the official CLI you already have signed in, speaks its streaming protocol, and leaves its sign-in, context files and transcripts exactly where the CLI keeps them. You can keep using the terminal alongside it, on the same projects, on the same day.
What Helios adds is not capability but visibility and defaults. Most of what it shows you — permission modes, plan mode, a dollar cap, checkpoints — exists in the CLI as a flag or a slash command. In a terminal you reach for those when you remember to. In Helios they are on the screen, on by default, and the same across three providers.
Side by side
| What you are doing | In a terminal | In Helios |
|---|---|---|
| Running several sessions | One per terminal tab or tmux pane; you keep the map in your head. | Every project and session in one sidebar with a live status dot; a session that is waiting on you says so. |
| Seeing what the agent just did | Scrollback. Tool calls and their results interleave with the answer. | A transcript that streams markdown and code, with one card per tool call showing its result and elapsed time. |
| Approving a command or an edit | A prompt in the stream: the command, y/n. | A dialog with the full command or diff, why the agent wants it, and the rule it would remember if you say “always”. |
| Choosing how much to trust it | --permission-mode, per launch. | Ask, Accept edits, Auto, Bypass, Plan or Never ask — per conversation, one click, and read-only whatever you pick when the folder is your home directory. |
| Capping what a run can cost | --max-budget-usd if you pass it. | The same native cap, required rather than optional, shown next to the composer. GPT sessions get a shared token budget that covers their subagents; OpenRouter gets a fixed allowance and a bounded tool loop. A trip stops the work and keeps your draft. |
| Agreeing on a plan first | Plan mode, a slash command away. | The same plan mode, plus a Work: an objective, a definition of done and an acceptance checklist that belong to you and that the model’s own plan can never rewrite. |
| Undoing a bad turn | The CLI’s own rewind. | A snapshot of the working tree before every turn, and a rewind dialog. Two chats in one folder can both write; a clobber is recoverable rather than forbidden. |
| Picking a model | A model id you type. | A picker built from what the connected CLI actually ships, latest per family first — nothing hand-listed that the binary does not have. |
| Watching subagents | Nested output in the same stream. | An activity strip keyed on the delegation the agent actually made — an actor exists only when the CLI reports one, and never changes state on a guess. |
| Switching provider | A different CLI with different flags. | Claude, GPT and OpenRouter in one window with the same permission chip, the same budget seam and the same plan pane. |
When the terminal is the right answer
You are not at a desktop
Over SSH, on a server, in a container, on a machine without GTK 4: the CLI is the tool. Helios is a native Ubuntu application and does not pretend otherwise.
You are scripting it
claude -p in a pipeline, a Makefile, a CI job. Non-interactive runs have no use for a window.
You run one session and it is short
If you rarely have more than one agent going and rarely run anything long enough to cost real money, the visibility Helios adds is answering a question you are not asking.
When Helios earns its place
Several agents at once
Three projects, five sessions, one of them waiting for a permission you have not noticed. The sidebar is the map you were keeping in your head.
Long autonomous runs
An hour of tool calls on per-token billing is where a cap stops being a flag you might set and becomes the thing standing between you and a surprise. Helios will not start a Claude session without one.
Work you have to answer for
When someone will ask what the agent did and why you let it: the transcript with one card per tool call, the permission dialogs with the full diff, and the Work whose definition of done you wrote are the record.
What you give up
Honestly: a little. Helios adds a process between you and the CLI, so a brand-new CLI feature may take a release to surface in the picker or the dialogs, and there is a desktop app to install and update. The tradeoff is deliberate — every control it shows is one the CLI already has, so nothing you learned in the terminal stops being true.
What you do not give up: your sign-in, your CLAUDE.md and memory files, your transcript history, or the terminal itself. Helios reads and writes the same ~/.claude/projects the CLI does, keeps its own state under ~/.helios, and sends nothing anywhere on its own.
Next
- Helios — what it does, install, where your data lives
- User guide — first run, permissions, budgets, Works and plans
- Install on Ubuntu 26.04 — one APT repository,
sudo apt install helios