Installing and Launching Agents
Install and launch agent CLIs in Daintree: availability checks, launch-gate recovery, the shared launcher, pin and shortcut controls, first-agent setup, persistent launch flags, screen mode, and permission bypass.
Installing Agents
Each agent is a separate CLI you install yourself. Install and authenticate one and Daintree picks it up on its next availability pass, with no configuration on your side.
Settings > CLI agents opens on its General page, whose Agents section lists the agents that need attention first, then folds the ready ones and the ones that aren't installed behind Show toggles. Re-check runs the availability probe again, and Run setup wizard walks you through installing one. Pick an agent from the picker at the top to open its own page. Under Launching, Pin to toolbar controls whether its button is shown; turning it off still leaves the agent available in the launcher.
An agent's page leads with its install status whenever the agent is not ready: Not installed, Needs setup, Blocked, or No credentials detected. That section shows where the probe found the binary (Detected CLI), the install commands for your OS when it is missing, the agent's troubleshooting notes, and a link to its official documentation. Agents that publish a usage page get a View usage button beside the picker. For Codex, an Account quota section reads the signed-in account's limit windows from Codex itself (for example a weekly limit and when it resets); Codex only reports one when you sign in with ChatGPT rather than an API key. Claude Code does not share usage with other apps, so its page points you to View usage instead.
Install commands, authentication steps, and resume specifics live on each agent's own page, linked from the agent roster.
Launching an Agent
There is one searchable launcher. It is the same surface whether you open it from the toolbar or from the dock's + button: same inventory, same rows, same keyboard model.
Type to filter across agents, panels, and recipes at once; search results are one flat list, each row naming what it is (Agent, Panel, or Recipe). Unfiltered, the launcher lays everything out in two columns, folding to one when it is narrow:
- Agents: one column holding every agent. The agents you launched most recently come first, marked Recent, then the rest with pinned agents at the top. Each agent is listed once. Installed agents that are not launchable yet sit at the bottom of the column marked Setup: they are offered, not hidden, so there is always a route to finishing setup. When no agent is installed at all, the column lists Available agents with a Set up agents row.
- Open in dock and Open in grid: panel kinds, grouped by where they will actually land rather than hidden when they cannot dock.
- Recipes: your recipes, with their scope shown on the right, or a Create a recipe row when there are none.
- A More footer with Manage agents and Customize toolbar.
Each row carries the affordances that used to be scattered across trays and submenus:
- Pinning: press Alt+P on the selected row, or click its pin button, to add or remove its toolbar button. Agents, panels, recipes, resumable sessions, and plugin launch entries can be pinned. Pins retain the specific launch target: pinning an agent preset does not change the worktree's default preset, and recipe pins distinguish their project or plugin origin.
- Shortcut capture: press F2 on a built-in agent's row, or click its keyboard button, to record a launch shortcut in place, without going to Settings.
- Preset rows: an agent with presets names the preset a plain Enter will launch, and Right Arrow expands it into sibling rows: Default first, then CCR routes, Project shared, and Custom, with the current one marked Current. Left Arrow closes them. They are siblings rather than a submenu so arrow-key navigation stays in one list.
- Attention pips: a dot on an agent's icon means one of its sessions in the active worktree wants you: amber when one is waiting, blue when you have an unsent prompt in one (directing). Agents that are only working, finished or idle get no dot, so the few that need you stand out.
Agents also launch straight from a bound keyboard shortcut, from the agent buttons on a worktree card, and from the toolbar buttons of pinned agents. In a Git project, these ordinary launch controls use the active worktree's directory and resolved preset. A lightweight workspace uses its workspace folder instead.
Starting an agent with a new worktree
The new worktree dialog can start an agent for you once the worktree is ready. Pick one in its Agent field and, optionally, give it a first task in Prompt. Daintree creates the worktree, waits for its setup to report ready (up to five minutes), then launches the agent with that prompt as its first turn. The dialog remembers the agent you picked per project.
If setup fails, times out, is still running, or is waiting for you to approve its commands, the agent is not started into an unprepared tree. A notice above the grid, Agent not started, says why, keeps your prompt, and offers Start with the agent's name to try again. A recipe or a cloned layout that already starts an agent takes precedence: the dialog warns you when the selected recipe does, and the same notice explains it afterwards.
Launch flags survive restarts
Flags supplied by whatever launched an agent, such as the Daintree Assistant or an MCP client, are recorded with the panel and re-applied after Daintree rebuilds the command: on restart and resume, on a cold restore, when a panel is duplicated, when a preset falls back, and when the launch gate continues. A multi-token flag such as Codex's -c key=value stays paired.
When an Agent's CLI Isn't Available
Every agent resolves to one of five availability states. Daintree probes in layers: PATH first, then the agent's known native install paths, then package-manager probes such as npm-global or known Python-tool bin locations, then WSL on Windows for agents that only ship a Linux binary. On macOS and Linux, it checks executable files directly in PATH order and treats symbolic links to the same binary as one install; it does not require an external which command. Windows uses where.exe. A separate passive authentication probe runs alongside it, so installed-but-not-signed-in is distinguishable from installed-and-ready. See the Linux release changes.
If the login-shell PATH refresh fails on Linux, Daintree adds existing common install directories, including a user npm prefix, Cargo, Bun, Deno, Linuxbrew, ~/bin and /snap/bin. These Linux-specific additions go after the PATH it already has, so they do not replace a binary your PATH already finds. Linux prerequisite instructions also follow the distribution's package-manager family; commands needing sudo remain manual, copyable instructions.
| State | Meaning |
|---|---|
| Missing | No probe found the binary. Shown as Not installed. |
| Installed | Binary found, but not directly launchable from the host, typically because it was detected only inside WSL on Windows. Shown as Needs setup. |
| Ready | Binary found and launchable; the local authentication probe found supporting evidence. |
| Blocked | Binary exists, but execution was denied. The reason is reported as security or permissions. |
| Unauthenticated | Binary found and launchable, but the auth probe found no credentials. Shown as No credentials detected. Still launchable: the CLI may already work, or ask you to sign in itself. |
Ready and Unauthenticated both count as launchable. Anything except Missing counts as installed.
The launch gate
Selecting an agent whose CLI is not launchable does not deep-link you into Settings. It opens a launch gate in the panel the agent would have occupied: the launch is held, not cancelled. The gate shows:
- A diagnostic banner naming the actual state: binary not found, detected in WSL, blocked by security software, or sign-in not detected. Where Daintree has one, it prints the last known path.
- The install commands for your OS, taken from the agent's registry entry, each one copyable, plus that agent's troubleshooting notes.
- Re-check, which forces an availability probe past its usual throttle. If the CLI has become launchable, the gate continues the launch it stood in for, resolving the command against the path the probe just found.
- Run anyway, which proceeds with the original launch regardless. Use it when you know the binary works and the probe is wrong: a wrapper script, an unusual shell environment.
- Agent settings and a Docs link, as secondary routes.
Unlaunchable rows in the launcher say so before you select them: they are marked Setup, and their tooltip reads "… needs setup. Select to see recovery options", or "… is blocked by endpoint security. Select to see recovery options".
chmod +x on the path Daintree shows. Use the reported path and reason to fix that execution restriction before trying again.Alt-Screen vs Inline Mode
A terminal CLI can draw itself two ways. Alt-screen mode takes over the alternate screen buffer and repaints a full-screen TUI, which is what the CLI does natively. Inline mode writes ordinary lines into Daintree's own scrollback, which keeps WebGL scrollback, search, and clean resize behavior. Neither is universally better; inline is smoother inside a tiled panel, alt-screen matches what the CLI looks like in your own terminal.
The decision resolves from three levels, most specific first: a preset override, then the agent's own setting, then the global Use alt-screen mode switch in Settings > CLI agents > General > All agents. The per-agent and per-preset Alt-screen mode control is tri-state (Default, Inline, or Alt screen, with the Default option naming what it currently resolves to), so an explicit choice at a narrower scope always beats a broader one.
The global switch is off by default, and no shipped agent pins a curated default of its own, so every untouched agent resolves to "inline". What that means in practice depends on whether the CLI has a flag for it:
- OpenCode (
--mini), Codex (--no-alt-screen), and Grok (--no-alt-screen) declare an inline flag, so they launch inline. - Agents without an inline flag keep their CLI's rendering behavior, subject to Daintree's per-agent terminal handling. Gemini, Antigravity, Qwen, and Mistral Vibe block alternate-screen escapes so output stays in scrollback; other agents may use their own full-screen UI. A missing flag does not imply every CLI renders the same way.
- Grok is the one agent that also declares the opposite flag: choosing alt-screen for Grok passes
--fullscreen, forcing full-screen rather than merely allowing it.
The alt-screen control only appears in Settings for agents that declare a flag in one direction or the other. For everything else the setting would have nothing to do.
Permission Bypass
Most agent CLIs ask before they edit a file, run a command, or reach the network. Some provide a supported flag that turns those prompts off; others do not. Daintree can inject that flag for you, and treats it as the high-blast-radius setting it is.
Settings > CLI agents > General > All agents > Skip permission prompts is the global switch. It defaults off. It is a live override, not a stored default: it is evaluated at flag-generation time on every spawn, restart, restore, and resume, and it never rewrites your per-agent choices. Turn it off and it stops contributing bypass on subsequent launches, including resumes captured while it was on. An explicit per-agent or preset Skip all still applies; running processes keep the flags they started with. It applies to agent terminals only; Daintree Assistant sessions carry their own separate setting.
Per agent, and per preset, the Permission prompts control has up to four choices:
- Inherit: defer to the next level up. The option names what it currently resolves to, such as Inherit (CLI default) or Inherit (Skip all).
- CLI default: Daintree adds nothing, and the CLI asks as it normally would. A true veto: a preset set to CLI default beats an agent set to Skip all, and an agent set to CLI default beats the global switch. The narrower scope always wins, so an explicit denial can never be overridden from further out.
- Auto, only for agents that declare an auto mode, which today means Claude Code (
--permission-mode auto) and Codex (--approve-for-me): the agent approves actions under its own rules instead of asking about each one. For an agent without one, a saved Auto shows disabled and launches with the CLI default. - Skip all: pass the agent's bypass flag for this scope, approving every file, command and network action.
The row shows the arguments the resolved choice adds to the command. The global switch only ever means Skip all, and it reaches every agent with a bypass flag Daintree knows (see the table below); OpenCode, GitHub Copilot CLI, Goose and Mistral Vibe have none, so they need their own custom arguments if you want one. In the agent picker, an agent that will launch with bypass on is marked Skips permission prompts.
The flag injected for each agent:
| Agent | Bypass flag |
|---|---|
| Claude Code | --dangerously-skip-permissions |
| Gemini CLI | --yolo |
| Antigravity | --dangerously-skip-permissions |
| Codex CLI | --dangerously-bypass-approvals-and-sandbox |
| Grok | --always-approve |
| Cursor | --force |
| Open Interpreter | --auto_run |
| Amp | --dangerously-allow-all |
| Aider | --yes-always |
| Qwen Code | --yolo |
| Kimi Code | --yolo |
| Crush | --yolo |
| Kiro CLI | --trust-all-tools |
| OpenCode, GitHub Copilot CLI, Goose, Mistral Vibe | None |
Four agents have no bypass flag at all, so the setting does nothing for them. OpenCode is the one worth calling out: it does have a skip-permissions flag, but only on its run subcommand, not on the interactive TUI launch Daintree uses. It is excluded deliberately rather than by oversight, so there is nothing to look for.
When a scope resolves to bypass-on, the Settings editor shows the exact flag it will append to the command, so the consequence is visible before you launch.