Vectrum Code

A native T3 Code-style desktop client for coding agents, built with GPUI and GPUI Kit. It drives Claude Code, Pi and jcode.

Every thread runs its own agent process in the background. Claude Code threads talk to the claude CLI over the stream-json protocol the Claude Agent SDKs use (--input-format/--output-format stream-json plus the stdio control channel). Pi threads talk to pi --mode rpc over its JSONL RPC protocol. jcode threads talk to jcode api --stdio, a bridge to jcode's versioned harness API that starts jcode's shared daemon if it isn't running, so jcode threads also show up in jcode's own session list.

The layout and workflow follow T3 Code. The look follows MetalUI, the "soft hardware" design system vectrum.cc is built on: the conversation sits on a raised sheet, buttons are caps that sink when pressed, code and tool output sit in inset wells, and labels are engraved in mono caps. It comes in MetalUI's two colorways, Bone and Graphite, plus popular color schemes rebuilt in the same style: Tokyo Night, Catppuccin, GitHub, One, Nord, Gruvbox, Rosé Pine, Solarized and Dracula.

Features

• A sidebar laid out like T3 Code's inbox: unsent drafts, pinned and active threads, then the Working, Snoozed and Settled shelves. Each card shows the project, its branch or worktree, running processes, the pull request, the agent, and a status (Working with a timer, Approval, Input, Limited, Failed, Woke, Done) or how long ago it was active; hovering swaps that for Snooze and Settle, and a hover card sums the thread up. Search covers titles and conversations (↑ ↓ Enter, Esc clears), and the folder button filters the sidebar to one project. Drag the sidebar's edge to resize it.
• Settle a thread to stop its process and file it under Settled. Its transcript stays, and the next message resumes the conversation. Settling, snoozing, unpinning, archiving and discarding a draft leave a "Settled 1 thread, Ctrl+Z to undo" notice for 5 seconds.
• T3 Code's thread menu, on a right click in the sidebar or a click on the title in the chat header: New thread on branch, Pin, Settle, Snooze (in 1 or 3 hours, this evening, tomorrow, next week, or a custom date, time or duration), Wake, Rename, Regenerate title, Mark unread, Filter by project, Auto-settle behavior, Copy (path, branch, thread ID), Project settings, Archive and Delete. Ctrl+click or Shift+click picks several threads for one menu; a double-click renames in place.
• A command palette (Ctrl+K) with T3 Code's actions (new thread in a project or without one, copy the thread ID or PR link, restart the agent, add a project, change the theme or appearance, usage, settings) and recent threads; typing also finds projects, settings and threads, and a leading > searches actions only.
• The legacy sidebar (Settings → Legacy sidebar): threads in per-project trees, with T3's sort menu (projects by last message, creation or by hand, threads by last message or creation, how many show) and project menu (Rename, Group into..., Copy Path, Project settings, Remove).
• A new thread opens on a start screen, "What should we build in project?", with the project as a picker (No project, the projects, Add project) and a way to start without a project ("What should we work on?").
• The composer is laid out like T3 Code's. Its footer holds the model, reasoning and access controls (they lose their labels, then fold into …, when the composer is narrow), the attach button and the send button, which turns into stop, steer or queue while the agent works. Its placeholder says what it takes: a message, your own answer to a question, feedback on a plan.
• Under a new thread's composer hangs T3's strip: the machine, the workspace (Current checkout, New worktree or Previous worktree) and the branch, with a searchable branch picker (current, worktree, remote and default branches, Create new ref, and Start from origin for a new worktree). Settings → Appearance → Composer context keeps it there once the thread has started.
• A model picker across providers (Ctrl+Shift+M): a rail with Favorites and each provider, Claude's models (current ones, plus legacy models folded away), every model Pi's configured providers offer, and every model route jcode's logins and provider profiles make available. Star models to keep them under Favorites, search across all of them, pick the first nine with Ctrl+1–9, and switch providers with Ctrl+Shift+↑/↓. Picking another provider's model starts a thread of that kind in its place.
• A reasoning menu (Ctrl+Shift+E): Low to Max (as the model supports, its default marked), Claude's Ultracode and Ultrathink, and Fast Mode on models that have it. Pi threads list Pi's thinking levels; jcode threads offer effort levels on its Anthropic, OpenAI and OpenRouter routes. All of it switches live.
• An access menu (Ctrl+Shift+A): Supervised, Auto-accept edits, Auto or Full access. Pi and jcode have no approval prompts, so their threads always run with full access. With plan mode turned on (Settings → General → Legacy features), a Build/Plan toggle and Shift+Tab switch Claude into plan mode and back.
• Finished turns fold their tool calls and interim messages into one Worked for 12s row above the answer. Click it to unfold.
• Under each turn that touched files, N changed files +a −r with a folder tree and per-file counts. Click a file, or Open diff, to read the diff in the app. The app snapshots the working tree through a private git index, so your staging area is never touched.
• Streaming Markdown with a copy button on every code block, and tool calls as collapsible cards with their input and output.
• What the agent waits on docks on top of the composer, as in T3 Code: an approval (the command or file, Decline, Approve, and Cancel or Always allow this session under …), Claude's questions (one at a time; pick with a click or 1–9, a single choice moves on by itself, tick several on multi-select ones, or type your own answer in the composer; Previous, Next question, Submit answers; fold or dismiss it), a Pi extension's confirm or choice, or Claude's plan (Plan ready, then Implement, Implement in a new thread, or type feedback and Refine). The transcript keeps what you answered.
• Docked there too: queued messages (drag to reorder, edit in the composer with Alt+↑, send now as a steer, remove), work left running in the background ("Running: …" with Stop), a usage limit, and the agent's task list with its progress.
• Ctrl+S puts the prompt aside; a Stash tab beside the composer brings it back.
• Attachments: images are sent inline, other files by path. @ lists the thread folder's files and folders to mention. An optional ring (Settings → General → Legacy features) shows how full the context window is, with the numbers and Compact context on hover.
• A project panel: the machine, Open in your editor (Zed, VS Code, Cursor, VSCodium, Sublime Text, Kate or the file manager, whichever are installed), the branch with a switcher, Commit (asks the agent to commit, or commit and push), the size of the uncommitted changes, and the thread's cost.
• Background tasks: when a command, monitor or subagent Claude left running finishes, the transcript says so and Claude picks the work back up and reports back on its own. A long-running command can be sent to the background from its tool card, like Ctrl+B in Claude Code.
• A browser the agent and you share (Ctrl+Shift+E, or the globe in the top bar), on Vercel's agent-browser: each thread has its own headless Chrome session, which its agent drives with the agent-browser CLI while the panel shows it live. Click, scroll and type in the page yourself, or go to a URL. When the agent opens a page while the panel is closed, a light on the globe says so.
• Interrupt with Esc or the stop button. Messages sent while the agent works are queued.
• Crash recovery: if a process dies, the next message restarts it and resumes the session.
• Threads are saved as you go and come back after a restart, settled ones included; the next message resumes the conversation. Optionally, threads a crash or restart interrupted continue on their own.
• New threads can start in a git worktree of their own, made like T3 Code's when the first message is sent: a temporary vectrum/<8 hex> branch checked out in a folder named after it (under ~/.local/share/vectrum-code/worktrees/<project>, or the worktree location in Settings → Storage), optionally from origin's copy of the base branch, with submodules filled in. A setup card shows each stage (with Work locally and Cancel), and the branch is then renamed after the first message (Settings → Source Control → Worktree branch naming: a prefix plus a few words, a semantic prefix like feat/, or your own instructions). Ctrl+Shift+L starts in the project's previous worktree instead.
• Thread titles written by a small model from the first message (Settings → Text generation).
• Pin threads to the top, snooze them until later, rename them, archive them to Settings → Archive. Working threads can fold into a Working shelf until they need you.
• When a usage limit stops a thread, it can snooze until the limit resets and continue then. Idle threads, and worktree threads whose pull request merged, can settle on their own.
• Send while the agent works to steer it, or queue the message for after the turn (and Ctrl+Enter for the other one). Queued messages wait above the composer.
• A / menu with Claude's commands and skills ($ for skills alone), docked on the composer.
• Diffs with hidden whitespace changes, foldable files, line numbers, and a stacked or side-by-side layout.
• Remote connections, like T3 Code's environments: turn on Allow remote connections on one device, and paste its pairing code (address and key) under Settings → Connections on another. The host's projects and threads appear in the sidebar tagged with its name; open, continue or start threads there and they run on the host. The connection is a WebSocket guarded by the key but not encrypted, so use it on trusted networks or over a VPN.
• A settings view (Ctrl+, or Settings in the sidebar), laid out like T3 Code's:
  • General: what new threads start with (model, reasoning, permissions, workspace, submodules), the text generation model, projects and threads, organization of the sidebar, and confirmations (unpin, settle, delete, and the quit shortcut).
  • Behavior: notifications (desktop, with sound, in-app), time format, response streaming, turn folding and costs, proactive diffs, diff defaults, the composer (skills in the / menu, collapsing on scroll, send key, follow-up behavior), update checks, continuing after restarts, background activity, and the editor "Open in" uses.
  • Appearance: comfortable or compact thread cards.
  • Themes: Light, Dark or System mode (System follows the desktop), and a light and a dark theme picked from live previews: Bone and Graphite; Tokyo Night (Night, Storm, Day); Catppuccin (Mocha, Macchiato, Frappé, Latte); GitHub (Dark, Dark Dimmed, Light); One Dark and One Light; Nord; Gruvbox Dark and Light; Rosé Pine (Main, Moon, Dawn); Solarized Dark and Light; and Dracula.
  • Providers: every agent instance (several Claude, Pi or jcode instances side by side, say two accounts), each with a switch, its version, login and plan, display name, binary, config folder, launch arguments and environment variables, plus updates and the health check interval, and usage hubs (CLIProxyAPI servers, with their management key).
  • Usage: cost, tokens and plan limits over the past day, week, month or quarter, read from Claude Code's and Pi's local logs and any usage hubs, by provider, model and day.
  • Connections, Keybindings, an Archive of settled threads, and About.

Requirements

• Rust (edition 2024)
• The claude CLI, logged in (claude once interactively)
• Optional: Pi for Pi threads (npm install -g @earendil-works/pi-coding-agent, then log in to a provider with pi)
• Optional: jcode for jcode threads (curl -fsSL https://jcode.sh/install | bash, then jcode login)
• Optional: agent-browser for the browser panel (npm install -g agent-browser && agent-browser install)
• GPUI's Linux system libraries: Vulkan, Wayland/X11, xkbcommon, fontconfig, freetype (on Arch/CachyOS: vulkan-icd-loader plus your GPU's Vulkan driver, wayland, libxkbcommon-x11, libxcb, fontconfig, freetype2)

Run

cargo run --release -- /path/to/project   # defaults to the current directory

The shortcuts follow T3 Code's default keymap.

Shortcut	Action
Enter / Shift+Enter	Send / new line
Esc	Interrupt Claude
Ctrl+K	Command palette
Ctrl+N (or Ctrl+Shift+O)	New thread (picks the project when there are several)
Ctrl+Shift+N	New thread in the current project
Ctrl+Alt+N	New thread without a project
Ctrl+Shift+[ / Ctrl+Shift+]	Previous / next thread
Ctrl+1–9	Jump to a sidebar thread (hold Ctrl to see the numbers)
Ctrl+Shift+S	Settle the thread, or un-settle it
Ctrl+Shift+P	Pin or unpin the thread
Ctrl+Z	Undo the last settle, snooze, unpin or archive (while the notice shows)
Ctrl+Shift+C	Copy the thread's PR link, or its ID
Ctrl+U	Usage
Ctrl+Alt+A / Ctrl+Alt+Shift+A	Change the theme / cycle System, Light and Dark
Ctrl+Shift+M / E / A	Model picker / reasoning / access
Ctrl+Shift+G / X / H	Branch / workspace / machine (new thread)
Ctrl+Shift+L	Start in the previous worktree
Ctrl+S	Stash the prompt, or open the stash
Alt+↑	Edit the last queued message
Ctrl+F	Find in thread
Ctrl+B	Show or hide the sidebar
Ctrl+Alt+B	Show or hide the right panel (the browser)
Ctrl+Shift+B	Show or hide the thread details (project panel)
Ctrl+O	Add a project folder
Ctrl+,	Open or close settings
Ctrl+Q	Quit

Environment

• VECTRUM_CODE_CLAUDE_BIN: path to the claude binary. Otherwise it's looked up on PATH, then ~/.local/bin and ~/.claude/local.
• VECTRUM_CODE_PI_BIN: path to the pi binary. Otherwise it's looked up on PATH, then in ~/.local/bin, ~/.npm-global/bin, ~/.bun/bin, ~/.volta/bin and nvm's Node versions.
• VECTRUM_CODE_JCODE_BIN: path to the jcode binary. Otherwise it's looked up on PATH, then in ~/.local/bin and ~/.cargo/bin.

Favorite models and everything in the settings view are saved in ~/.config/vectrum-code/settings.json. The environment variables above win over the binary paths set there for each agent's first instance. Threads and the project list live in ~/.local/share/vectrum-code (threads/, state.json); threads started without a project run in its scratch folder.

Layout

• src/claude.rs: spawns and owns the agent process, claude, pi or jcode api --stdio (stdin writer and stdout/stderr reader threads).
• src/session.rs: per-thread state. Folds Claude's stream-json, Pi's RPC and jcode's harness API events into chat items, handles permission requests, control requests and extension dialogs, and snapshots each turn's changes.
• src/models.rs: the model catalog across providers, and favorites.
• src/git.rs: branch and change stats, working-tree snapshots, and diffs.
• src/settings.rs: the saved settings file, cached in memory.
• src/ui/chat.rs: start screen, transcript (with turn folding), tool cards, permission bar and composer.
• src/ui/pickers.rs: the model, reasoning and access pickers.
• src/ui/settings.rs: the settings view and its pages.
• src/ui/changes.rs: the per-turn change overview and the diff view.
• src/ui/workspace.rs: window root: thread sidebar, top bar and project panel, plus the git and editor helpers behind it. Its submodules hold the sidebar (sidebar.rs, and the legacy layout in sidebar/legacy.rs), the thread menus, snoozing, confirmations and undo (actions.rs), the chat header (header.rs) and the command palette (palette.rs).
• src/ui/icons.rs: the extra Lucide icons the app embeds.
• src/ui/metal.rs: the MetalUI layer. Colorway tokens and shadow recipes copied from MetalUI's CSS, the theme mapping for the stock widgets, and the primitives (caps, wells, keycaps, status LEDs, segmented switch).
• src/ui/themes.rs: the color schemes, the palette builder that turns each into a full MetalUI palette, the light/dark/system selection, and the Themes page.
• assets/fonts/: Geist, Geist Mono and Geist Pixel, embedded in the binary (SIL Open Font License, see assets/fonts/OFL.txt).

Not yet supported

There is no UI for Pi extensions that ask for free text, code blocks have no syntax highlighting, and T3 Code's project scripts and terminal drawer aren't here yet. jcode models can't write thread titles (every jcode run is saved as a session), and jcode's side panel, applets and background-task progress aren't shown.