Daily use

Sessions, the session tree, models, agents, tools, and every slash command worth knowing.

Starting loop

loop                              # interactive TUI in the current directory
loop --cwd ~/code/other           # somewhere else
loop --model anthropic/claude-sonnet-4-6
loop --session <id>               # resume a specific session

loop is directory-scoped: sessions belong to the folder you started in, and the last model you used in a folder is remembered per project.

One-shot mode

loop run "explain this repo"
echo "summarize the diff" | loop run -
loop run "fix the failing test" --max-steps 20

Prints the final response to stdout and exits. No UI, pipeable, safe in scripts. --max-steps caps how much work it will do.

Sessions

From the shell, loop sessions lists sessions in the current directory with ids, models, and timestamps.

Messages and slash commands typed while the agent is generating queue up and run when the turn ends. Esc interrupts.

Sessions are trees

This is the part that isn't like other agents. Every entry has a parent, so a session is a tree rather than a flat transcript — you can go back to any earlier message and take a different path without losing the first one.

• /tree — navigate the whole session as an ASCII tree. Fold branches, filter (no-tools / user-only / labeled / all), type to search, bookmark entries with shift+l. Selecting an earlier point branches there, and loop offers to summarize the abandoned branch — optionally with your own prompt — so its context survives the switch.
• /fork — pick a previous user message; the path up to it is copied into a new session, and the message text comes back to your editor for editing.
• /clone — duplicate the current conversation into a new session.

The practical use: try an approach, watch it not work out, branch back to before you suggested it, and try the other one — with the agent still knowing what went wrong the first time.

Models

Thinking level maps to whatever each provider calls it — Anthropic budget tokens, OpenAI/xAI/OpenRouter reasoning effort, Google thinkingConfig — so one setting works everywhere.

/scoped-models is how you stop paying frontier prices for grunt work: pin a cheap fast model to subagents and keep the expensive one for the main loop.

Agents

Tab on an empty prompt cycles the built-ins:

• default — the full toolset
• plan — read-only investigator. Its bash runs in a kernel-enforced read-only, network-denied sandbox, and is refused outright if that can't be enforced. It treats a plan as a deliverable: when the plan is final it calls the plan tool, which ends the turn and renders the plan as markdown for you to hand off or keep refining.
• data-analyst — SQL-first work against a /datasource

/agents builds your own: custom system prompt, tool allowlist, optional pinned model. It registers as /<name> for one-shot use — /reviewer look at the auth changes.

Subagents: the task tool forks the current agent into a fresh context window. Activity streams live in the task box while it runs, but only the final report enters your context — that's the point, a subagent can read fifty files and cost you one paragraph.

Tools

read · bash · edit · write · grep · find · ls · sql · task

Colored diffs, syntax-highlighted output, file previews. read also fetches http(s):// URLs as readable text, and takes offset/limit for large files. edit and write enforce read-before-modify within a session, so the agent can't overwrite a file it hasn't looked at.

Opt-in, via /settings:

• websearch — DuckDuckGo, no API key
• ask — lets the agent pause and ask you a multiple-choice question instead of guessing

Images

Paste with Ctrl+V, /attach <path>, or Ctrl+I for a file picker. Screenshot a broken UI or a stack trace and ask about it directly.

On macOS, Cmd+V cannot paste an image into a terminal: the terminal handles that keystroke and only knows how to type text. Ctrl+V is the one that works — it pastes an image when the clipboard holds one, and the clipboard's text when it doesn't.

Workspace context

• AGENTS.md / CLAUDE.md in the repo load automatically as workspace context. /init writes one for you by reading the codebase.
• *.md under ~/.loop/agent/skills/ or .loop/skills/ register as /skill:<name>, and the model can invoke them itself via the skill tool when a task matches the skill's description.
• *.md under ~/.loop/agent/prompts/ become slash commands.
• /memory keeps durable facts per project.

Cost

The footer shows live cost, usage, and context per step, subagents included. /cost breaks it down by session, directory, today, last 7 days, month, and lifetime per provider. Anthropic prompt caching is handled automatically, with cache breakpoints moved across multi-step turns.

loop cost audit reconciles the ledger against the transcripts if you want to check the numbers.

Every slash command

/hotkeys lists key bindings. /alias makes your own shortcuts.

Other surfaces

loop serve                # web UI + WebSocket RPC (opt-in via /settings, token-locked)
loop rpc                  # JSON-RPC over stdio
loop rpc --socket         # …over a Unix socket
loop man                  # the manual (--install writes it to the manpath)
loop completion zsh       # tab completion for bash, zsh, or fish

Next: automate it or drive it from your phone.