Claude Code Sessions

One session, one thread of work · desktop + mobile app · not the terminal · Sep 2026

The thing you noticed

Someone online was using a kanban-style board to route ideas into sessions. You then spotted your own app's status filter (Needs input · Ready for review · Working · Completed · Archived) and wondered if you were missing a feature. You're not — that filter IS the kanban board. It's native. What's actually off isn't the tool, it's how the sessions underneath it are scoped: your list — ! Web, General, General Coding, ! Main — reads as four standing project folders, not discrete pieces of work. That's the thing worth fixing.

What a session actually is

The five states in your filter

StateWhat it means
WorkingClaude is actively running tools or generating a response right now
Needs inputBlocked on you — a question, a permission prompt, a decision only you can make
Ready for reviewClaude finished a turn and is waiting for you to look at the result before it continues
CompletedThe task was signaled done — often after a verification step, or because you said so
ArchivedPulled out of the active list — done-done, or parked

This is effectively a live kanban board of every thread of work you have going, for free, with no board to maintain by hand. It only tells you something useful, though, if each card (session) is one coherent thing — five "General Coding" sessions all sitting in "Working" tells you nothing.

Local vs. cloud sessions — the icon next to each one

IconMeaning
💻 LaptopRuns on this Mac — filesystem, tools, MCP connections, and credentials all stay local. Closing the laptop pauses it
☁️ CloudRuns on Anthropic's infrastructure — keeps going after you close the laptop or lock your phone

Devices (the section with "Add device") is a separate concept from either of these — it's Remote Control: pairing your phone or another machine so it can steer a session that's actually running back on this Mac. It doesn't change where the session runs or how it's scoped; it just lets more than one screen watch and drive the same conversation.

When to start a new session vs. keep going

SituationDo this
Starting a genuinely new task, even in a familiar projectNew session, named for the task
Picking the same task back up hours or days laterResume that session — its context is still there
Pivoting to something unrelated mid-conversation/clear if you're staying, or just start fresh — don't just keep typing
Need to research or explore without cluttering the main threadA subagent — its findings come back as a summary, the digging doesn't
Something that should outlive this one conversation (a preference, a fact, a decision)Memory — not "keeping the session open so it remembers"
The one-sentence test: "Could I name this session for the one thing it's doing?" If the honest name is a project folder — ! Main, General Coding — it's a holding tank. If it's a sentence — Systems folder: Tailscale routing + workbench build — it's scoped right. (This very conversation drifted from a git-connector question into a folder reorg, a live-server build, and now this document — a real, lived-in example of exactly the pattern this page is about.)

Page 2 — Terms sheet

The words you'll hit using the app day to day, one line each.

TermPlain meaning
SessionOne conversation, one context window, tied to a working directory
Context windowEverything the model can currently "see" from this conversation — it has a limit, and fills up faster when a session covers many unrelated things
CompactionAutomatic summarizing of older parts of a long conversation so it can keep going past the context limit — lossy by nature, another reason not to let one session run forever across unrelated work
/clearWipe the current session's context and start clean, same window
SubagentA sub-conversation spun off to do research or a contained task; only its summary returns to the main session
Memory (auto memory)Facts, preferences, and project context saved outside any one session, so a brand-new session can still know them — this is the actual fix for "I need it to remember," not a long-lived session
Agent View / session listThe sidebar or claude agents view listing every session, grouped by status — the "kanban board" you were looking for
Working directoryThe project folder a session is rooted in — sets what files it can see by default, not what topic it's limited to
Local sessionRuns on your Mac; laptop icon; pauses when the laptop sleeps or closes
Cloud / background sessionRuns on Anthropic's infrastructure; cloud icon; keeps running after you close the laptop
Devices / Remote ControlPairing a phone or browser to watch and steer a session running elsewhere — access, not execution location
ResumeReopen a past session with its full context intact, to keep working the same thread
ArchiveTake a session out of the active list without deleting it — for done-done, or parked-for-now
Kitchen sink sessionThe documented anti-pattern this whole page is about: one session carrying many unrelated chunks of work
Slash commandA typed shortcut like /clear or /loop that triggers a built-in behavior instead of a normal message
HookA rule that runs a shell command automatically on some event (a tool call, a stop) — a terminal/config-level concept, not something the app surfaces directly
MCP server / connectorAn external tool or data source (Gmail, Calendar, a database) wired in so Claude can use it inside a session
Claude Code Sessions Primer · @ Research/@ Building · v1 · Sep 12, 2026 · Session-state and scoping guidance is from Anthropic's official Claude Code documentation (agent-view, desktop, mobile, remote-control, best-practices); the exact automatic triggers behind the mobile app's five-state grouping aren't fully documented publicly — cross-check the live labels in-app if they ever drift from this sheet.