Claude Code

Claude Code is Anthropic's interactive coding CLI. As a Subshell harness it is the fullest integration of the built-ins: the subshell mcp server is wired in at every launch with nothing to register, the CLI's own hooks drive your notification bell, and a restarted subshell resumes the same conversation rather than starting a fresh one.

Finding the binary

Subshell resolves claude on the machine that will run the pane: the claude on the PATH there (including what a login shell would find, which covers nvm- and volta-style installs), then Claude Code's known install locations under your home directory. If you keep it anywhere else, point at it explicitly with CLAUDE_PATH in that machine's environment. One host setting travels through to every launch when set: CLAUDE_CONFIG_DIR, so a non-default Claude config directory is respected.

Detection is the plane asking, the machine answering: the launch page and the Re-check button probe the binary, so the picker only offers Claude Code on hosts that actually have it. Nothing is installed on your behalf here: if detection finds nothing, Installing Agent CLIs covers the one-line installer, and the vendor setup docs the manual path.

What a launch looks like

Every Claude Code subshell starts as an interactive claude session in a tmux pane, with the command line built by the control plane: the subshell mcp registration first, then identity flags, then a --settings JSON carrying your preset's configuration, then the subshell's name via --name, then any preset and one-off flags. Settings ride on the command line per invocation: Subshell never writes into your real ~/.claude files, so a preset cannot leak into the Claude Code you run outside Subshell.

In the preset editor, Claude Code's known settings (permission mode, model, output style, and friends) appear as editable fields rather than raw flags, and the editor suggests the environment variables and flags that matter for this CLI. If a preset sets its own hooks key, Subshell's own hooks win over it; the alternative would be no notifications at all.

Cross-subshell tools, automatically

The subshell MCP server (the tools that let an agent see its siblings, read and write channels, and coordinate) is registered for every Claude Code subshell automatically. The control plane writes a per-subshell --mcp-config document holding just the subshell server entry, and the launch names it; from the first prompt the agent can call Subshell's tools.

Two things about that file are worth knowing:

• It contains no secrets. The config names only the command to run. The subshell mcp process starts as a child of the harness and inherits the credential baked into that pane's environment, each subshell's own scoped API key, so one machine's agents can never authenticate as another's.
• The same wiring lands on a node. Launch a Claude Code subshell on an enrolled machine and the registration travels with the launch, naming that machine's own subshell binary as the MCP server. Nothing on the node has to be configured.

Notifications and conversation identity

Claude Code is the one harness that exposes hooks, and Subshell uses them for two things:

• Attention. The Stop and Notification hooks report "this turn finished" and "this turn is waiting for you" to the control plane, which is what makes the notification bell and the waiting state work. Delivery is still gated by the subshell's bell and your account switch; a hook firing never bypasses them. See Notifications.
• Conversation identity. The SessionStart hook reports the transcript id whenever the in-pane session changes (/clear, /resume, /fork, /compact), so Subshell always knows which conversation to resume.

The hook commands are not guesses: the control plane resolves, per machine, the subshell report command line that actually exists on the host running the pane, and hands the plugin that. On a desktop install with nothing but the Subshell binary beside the pane, the hooks still reach home. If a launch ever cannot resolve that command, Subshell omits the hooks entirely rather than wiring a command the machine cannot run: one launch without notifications instead of an error on every turn.

Resume across restarts

When a Claude Code subshell starts, Subshell pins its conversation id (--session-id), and a restart, manual or automatic, brings the same conversation back with --resume. That pin survives /clear and in-pane /resume precisely because of the SessionStart hook above. This is unique among the built-ins: OpenCode, Codex, Hermes, and pi have no per-subshell conversation pin, and restarts there begin fresh.

See also

• Notifications: what the hooks feed, and how delivery is gated
• Presets: saved per-harness configuration, including Claude Code's settings fields
• The MCP server: what the registered tools actually give the agent
• Installing Agent CLIs: the in-page installer for the claude binary

Edit on GitHub