Your Claude Code sessions, in a browser tab.
LAST lists every Claude Code session on your machine and lets you keep chatting in any of them. It
reads the transcripts your terminal already writes and continues the session through the Claude Agent
SDK, so a reply you send here is waiting for you the next time you run claude --resume.
- Every session in one list. Grouped by project with git branch and relative time. Groups are collapsible (collapsed by default), a Recent shortcut keeps the latest few on top, and a project header opens that group's full session list in a modal.
- The whole conversation. Markdown, syntax-highlighted code, tool calls and thinking blocks, replayed the way the terminal showed them.
- Reply from the browser. Same folder, same session, same Claude. The turn is appended to the same transcript file your terminal uses.
- Live. Leave a session open in the browser, keep working in the terminal, and the page follows along on its own. No refresh.
- Streaming replies with token count, cost and duration for every turn.
- Permission prompts in the browser. Allow once, allow for the session, or deny.
- Clarifying questions as a proper dialog. When Claude asks a multiple-choice question, it opens as a modal with each option's explanation next to it, not just plain chat text.
- Permission modes in the composer: Ask before changes, Auto-accept edits, Plan only, Auto
(the default, as in Claude Code: Anthropic's own classifier approves or denies routine calls;
anything it isn't sure about still asks you), or Bypass permissions (the CLI's
--dangerously-skip-permissions: nothing asks). - Queue while Claude works, or send now. A message typed during a turn waits in a queue. Its
Send now button cuts in like
ctrl+x ctrl+sin the terminal: Claude stops what it's doing and answers the message straight away, keeping everything it had already done.⌘Jor⌘⇧↵sends what you typed that way without queueing it first. - Model and effort pickers, and a light / dark / system theme.
- Works on a phone. The sidebar becomes a drawer and the whole app resizes down to a small screen.
- Long replies as cards on a phone. Under a long answer, Show as cards splits it at its headings and bold labels and opens the pieces full screen, one card at a time. Swipe left and right to move between them.
- Search this session.
⌘Ffinds text in prompts, replies, thinking and tool calls of the open session, opens the card that holds each match, and steps through them newest first. - Keyboard first.
⌘Ksession palette,⌘Fsearch this session,⌘Bsidebar,⌘Jsend now,Escstops a running turn. - Nothing leaves your machine. No telemetry, no analytics, no account, no server-side database.
- Node.js 20.9 or newer
- macOS or Linux (Windows is untested)
- Claude Code installed and signed in. LAST reuses your existing CLI credentials, so there is no API key to set up
LAST talks to Claude through the copy of Claude Code bundled with @anthropic-ai/claude-agent-sdk,
not the claude on your PATH. The SDK is pinned to 0.3.310, which bundles Claude Code 2.1.310.
Newer Claude models need Claude Code 2.1.280 or newer, so if a reply fails with "Claude Code … does
not support this model", claude update won't fix it. The SDK version in package.json has to go
up.
npm ci
npm run devOpen http://127.0.0.1:4123.
You need one Claude Code session on disk before there is anything to show, so run claude in a
project once, then reload.
For a production-style run:
npm run build
npm startPress Ctrl-C to stop it. npm run check-bind tells you what is still listening on 4123.
There is no login and no token. Anything that can open a TCP connection to 127.0.0.1:4123 has your
whole Claude Code history and can run Bash on your machine as you.
So: do not run LAST on a shared or multi-user machine, and do not port-forward or tunnel port 4123. Check whether your editor forwards ports for you.
Every tool call asks you first in Ask before changes, Auto-accept edits, and Plan only. In Auto mode it doesn't: Anthropic's classifier silently approves calls it's confident about, and a call it's confident should be denied just shows up as a warning after the fact. Only the calls it isn't sure about still reach you. Auto is the default, so switch to Ask before changes if you want to see everything before it runs.
LAST loads no Claude Code settings files by default, so nothing on your CLI allow list runs
unprompted. The trade-off is that CLAUDE.md and AGENTS.md are not loaded either. Set
LAST_SETTING_SOURCES=project to opt back in, or user,project,local for full CLI parity.
Typing "ultracode" in a message is a built-in Claude Code trigger, not a LAST feature. It opts that turn into the Workflow tool, which can spawn several subagents at once. It needs Workflows enabled on your Anthropic plan and a model that supports the deepest effort level; LAST can't check either of those for you. LAST shows a line when a workflow starts and when it finishes, but not its live progress in between.
See CONTRIBUTING.md.
This is an unofficial, community-built project. It is not affiliated with, endorsed by or supported by Anthropic. "Claude" and "Claude Code" are trademarks of Anthropic. LAST is provided as is, with no warranty. You are responsible for what the agent does on your machine.
LAST depends on @anthropic-ai/claude-agent-sdk, which is proprietary to Anthropic PBC and is not
covered by LAST's licence. It is fetched from npm under Anthropic's own terms, and it spends your
Claude credits.
UI ideas were inspired by siteboon/claudecodeui. No code was copied from it.