The open-source agent IDE
Clave
Every agent session is a terminal. Clave gives the terminals a home: tabs, groups and whole workspaces you launch in one click, whichever agent you run.
How it works
Terminals, grouped the way the work is.
-
01 Sessions
Every agent session is a terminal. Clave gives it a tab.
Claude Code, Codex, Antigravity, Pi, or a plain shell: each runs in its own tab, in its own terminal, with the agent’s state shown on the tab so you see who is working, who is waiting for you, and who is done. One, two or a grid of them on screen at once.
-
02 Groups and workspaces
Tabs group by project, and a project raises in one click.
A group remembers its sessions, their folders, its saved terminals (a dev server, a board) and the prompt each session opens on. A workspace is a folder with a small .clave file; Clave finds every project’s file underneath it and offers each as a template. A teammate keeps their own copy and never touches yours.
-
03 The agent drives the app
The IDE builds itself around the work.
Every Claude Code tab is connected to an MCP server inside Clave, in the terminal and in the chat view. The agent can open a sibling session with a prompt, launch a whole group, pin a live page to it, open a file or a document beside itself, message another tab, and ask you for a secret it never gets to see. Our own tools drive Clave this way; yours can too.
-
04 What the agent never sees
The agent asks for the key. You are the one who types it.
A secret is pasted by you into a masked field and injected for one action, or written straight to the file it belongs in. The value never passes through the transcript, and the agent never reads it back. The same discipline holds over SSH, where a machine elsewhere is just one more location in the sidebar.
The workbench
One window: your projects on the left, your agents side by side, and the page they are building right there.
And because every Claude Code tab is connected to the app through MCP, the agent inside one of them can open the next tab, pin the live page beside it, and hand the work on.
What it does
The whole of it, as it ships today.
-
The bench
The window itself: tabs, groups, and the pages that belong beside them.
- Bring your own agent Claude Code, Codex, Antigravity, Pi or any command that runs in a shell, with named launch profiles for how each one starts.
- Groups and templates One group per project, remembered sessions and servers, a primed prompt, and every nested .clave file discovered as a template.
- Live views Pin a page to a group or a tab: a dashboard, a preview, a board. If its server is down, the view starts it.
-
Driven from inside
Everything a person can do to the window, an agent in a tab can do too.
- The MCP surface Twenty-three tools the agent inside a Claude Code tab can call, and the clave plugin with its skills, so the app is scriptable from within.
- The side panel A file tree with editors, an HTML preview, a Markdown page editor, and a git panel with diffs, log, worktrees and one-click sync.
-
Yours, and local
No account, no cloud backend, nothing between you and the model.
- Secrets, kept from the agent A secret is pasted by you into a masked field and injected for one action. The agent asks; it never reads the value.
- Remote work Sessions and files over SSH and SFTP, so a machine elsewhere is one more location in the sidebar.
- History and usage Every Claude transcript on the Mac searchable in one place, a message trail per tab, and usage meters per account.
With Exos
Clave is where the work happens. Exos is the rail it runs on.
In our own workspace, one Exos project is one Clave group, declared in that project’s record repository, and the group’s view is the project’s board. A delivery lane is a Clave tab; a wave is one coordinator tab opening one tab per lane and talking to them. Exos owns the tasks, the workstreams and the record; Clave knows nothing about Exos and is driven from outside. That separation is the point: either half works without the other.
Who it is for
Developers
Several agents, one workbench
If you already run more than one coding agent at a time, locally or over SSH, Clave is the window you were missing.
Teams
One file per project, yours to override
A default .clave in the repo gives everyone the same groups; a personal copy lets each person diverge without touching the others.
Builders of agent tooling
A desktop host you can drive
The MCP surface our own products use to open tabs, attach pages and hand work between agents is the same one yours get.
Where it came from
A plain observation, and what followed.
Clave began as a plain observation: every agent session is a terminal, and terminals deserve a better home than a window you lose.
So Clave is a desktop app that hosts terminals. Your agents, whichever you already use: Claude Code, Codex, anything that runs in a shell. Sessions group by project, whole workspaces launch in one click, and a live page can sit right beside the sessions producing it.
We work inside it all day, on everything. It is open source, and deliberately unaware of anything else we build: our own tools drive it from the outside, and yours can do the same.
Open source, all of it
Not a free tier around a paid core.
The entire app is open source under the MIT license, with no account, no cloud backend and no wrapper around the model. Whatever runs in a shell runs in Clave, and switching agents never means switching homes.
- clave.work The product’s own site.
- Download macOS, one universal app for Apple Silicon and Intel. An early Windows build exists.
- The repository Open source under the MIT license. Read it, fork it, wire your tools into it.
- Exos, the rail it runs on How the two fit: Clave is where the work happens, Exos keeps the record.