What is Lume?

Lume is a code editor with a coding agent at the centre of it. The agent can read your project, write and change files, run commands in a real terminal, and put every edit in front of you as a diff you accept or reject. It is built to cover the whole loop rather than one slice of it — writing something new, working out how an unfamiliar codebase fits together, chasing down a bug, and getting the change landed. The dial you actually tune is how much you hand over: keep your hands on the keyboard and let the agent assist, or hand it the goal and review what comes back.

Use cases

Three situations ask very different things of an editor, and Lume is shaped for all three.

• Starting something from scratch: describe what you want in plain language and let the agent scaffold the project, install what it needs, and run it. You review the result as a set of diffs instead of a wall of generated text.
• Working inside an existing codebase: the codebase index and # file mentions mean the agent starts from what your project actually contains, not from a guess. This is where most of the day goes, and it is what the context features exist for.
• Getting through volume: when a task is well-defined but tedious — a mechanical refactor, the same change across twenty files, a migration — SOLO mode lets the agent run it end to end while you check the output.

Dual modes

Lume ships two workspaces over the same engine, and Ctrl+Alt+\ flips between them.

• IDE mode: the editor is the main surface and the agent sits in a side panel. You drive, the agent assists. This suits incremental work in code you already know, where you want to see every keystroke land.
• SOLO mode: the layout turns task-first, with its own tool panels around the conversation — Doc, Terminal, Preview, Diff, History, and integrations. The agent drives, you define the goal and review the results. This suits larger self-contained tasks where following each step is less useful than checking the finished change.

AI coding

Four pieces do the AI work, and each one is separately configurable.

• Model: pick the model per conversation from the built-in list, or point Lume at your own provider endpoint and use your own key. Models that support a larger context window expose a Max Mode toggle on the model card.
• Agent: the agent plans, calls tools, and verifies its own work. It can spawn subagents to keep a side investigation out of the main conversation, and hooks let you run your own commands at fixed points in its loop.
• CUE: inline code completion as you type, driven from the status bar entry at the bottom right — accept with Tab, or snooze it for a while when you want quiet.
• Context: an index of your codebase, # to attach a specific file to a message, and automatic compaction so a long session does not simply fall off the end of the model's context window.

Development toolchain

Lume is a full editor, not a chat window bolted to a file tree.

• Editing and debugging: it is built on VS Code, so the editor, language support, breakpoints and tasks behave the way you already expect them to.
• Source control: the Git panel is there, and changes the agent makes flow through it like any other change — you stage, review and commit them yourself.
• Agent-powered code review: the agent can review a change set and report findings in the source control view, so a second pass does not depend on you remembering to ask for one.
• Extensions: extensions install from the Open VSX registry at open-vsx.org.
• Remote development: connect to a remote machine over SSH, or to a WSL distribution on Windows, and work there with the agent running against that filesystem.

Security

Two controls decide what leaves your machine and what runs on it.

• Privacy Mode: when Privacy Mode is enabled, Lume will not use your information like your chat interactions (including related code snippets and AI-generated outputs) for analytics, product improvement, and model training. The toggle is in Settings → Account → Privacy Mode, and it is off by default.
• Command execution: every command the agent wants to run passes through an auto-run mode that you choose — from asking you each time, through running inside a sandbox or against an allowlist, to running everything automatically. A red list of destructive commands always asks for confirmation regardless of the mode.