neuron-explorer
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
Vorn - The agent command center
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
Run agents interactively or let them work autonomously — everything local, everything yours.
Vorn is the agent command center. How you work is up to you.
Like vibecoding interactively? Open multiple terminals and pair with several agents at once — Claude on one task, Copilot on another, Codex on a third. Every agent runs in its own PTY with full native output. No wrappers, no API keys, no reimplementation.README 전문 보기 ↗Vorn is the agent command center. How you work is up to you.
Like vibecoding interactively? Open multiple terminals and pair with several agents at once — Claude on one task, Copilot on another, Codex on a third. Every agent runs in its own PTY with full native output. No wrappers, no API keys, no reimplementation.
Prefer to stay hands-off? Define your tasks and workflows, then let agents work headlessly in the background. Set up a workflow that automatically reviews code when a task moves to "in review", spins up another agent to generate documentation, runs your test suite after every implementation.
Chain agents together. A workflow can launch an agent to implement a feature, then trigger a second agent to review the diff, then run a script to deploy, all on a schedule or triggered by task status changes. Everything happens on your machine, orchestrated by you.
MCP-powered. Agents can read and write tasks, trigger workflows, and query project state through the built-in MCP server, no extra configuration needed.
Early stages. Vorn is in alpha. The focus right now is stabilizing the core and expanding from there. On the roadmap: SSH remote sessions, shared workspaces for team collaboration, connectors to pull tasks from GitHub Issues, Linear, Jira and more, and a companion web/mobile app so you can monitor your workflows and connect remotely while away from your computer.
macOS / Linux:
curl -fsSL https://raw.githubusercontent.com/vorn-run/vorn/main/install.sh | sh
Windows (PowerShell):
irm https://raw.githubusercontent.com/vorn-run/vorn/main/install.ps1 | iex
Homebrew (macOS):
brew tap vorn-run/tap && brew install --cask vorn
Or download directly from GitHub Releases.
Each release ships Vorn, the desktop app built on Vorn's GPU UI:
| Platform | Download |
|---|---|
| macOS | Vorn-<version>-arm64.dmg or -x64.dmg (signed and notarized) |
| Windows | Vorn-Setup-<version>.exe, a per-user installer |
| Linux | Vorn-<version>.AppImage or vorn_<version>_amd64.deb; needs WebKitGTK 4.1 and GTK 3 |
On Linux ARM64 boards, install.sh installs the server alone from
vorn-server-<version>-linux-arm64.tar.gz: vornd, its session holder and the vorn command in
~/.local/lib/vorn, run by the systemd user unit vornd. Connect to it from another device. Its
options, such as --host 0.0.0.0, go in ~/.config/vorn/server.env; VORN_NO_SERVICE=1 installs
without starting it.
The app updates itself from the release feed for the channel chosen in Settings (stable or beta): it downloads in the background, checks the checksum and signature, and swaps itself in when you restart. Terminals survive the restart on macOS and Linux. The .deb does not update itself; Vorn says when a new one is out. Windows builds are signed only when the release has a code-signing certificate, so SmartScreen may ask before the first run.
Installing Vorn also puts a vorn command on your PATH. It talks to the same server the app
does, and starts one if none is running. The Windows installer adds its folder to your user
PATH, and the .deb installs /usr/bin/vorn. On macOS and from the AppImage, the app links it
on first launch, into /usr/local/bin when you can write there and ~/.local/bin otherwise,
and says where. If that folder is not on your PATH, the notice offers the fix.
vorn session start --agent claude --prompt "fix the failing test"
vorn session list
vorn session read <target>
vorn session send <target> "run the tests"
vorn session wait <target> --timeout 10m
vorn events --json
vorn layout open dev
vorn workflow list
vorn workflow run "Nightly review" --input pr=42
vorn workflow stop <run>
vorn workflow check .vorn/workflows/*.toml
vorn server serve # a server on its own, without the app
vorn on its own opens the app. --json gives machine-readable output on stdout, with
notices and errors on stderr, so a command can be piped. Settings → General → Command-Line
Tool shows where the command is, and installs or removes it.
Every command that takes a session reads its <target> the same way, and the first rule that
matches decides:
3 is the third session in vorn session list;A rule that matches several sessions is ambiguous. The command lists the candidates on stderr
(with --json, on stdout as {"error":"ambiguous","candidates":[...]}) and exits 3. A command
that creates something prints its new id, and --json returns it.
| Code | Meaning |
|---|---|
| 0 | Done |
| 1 | Failed |
| 2 | The command line was wrong |
| 3 | A target matched more than one session or layout |
| 4 | No server could be reached or started |
| 124 | session wait --timeout passed first |
vorn session wait <target> blocks until the session's program exits and exits with its code;
--json prints the exit record. --timeout 30s (also 5m, 1500ms; a bare number is seconds)
gives up with 124. --status idle|done|blocked waits for the program's status instead, as the
app shows it (the program's own report, then hooks, then the screen rules). It waits on the
server's events and never polls.
vorn session read <target> prints the output the server kept. --lines N keeps the last N
lines. --scrollback, --unwrap and --format vt|html read the terminal itself, with its
soft-wrapped rows joined: vt keeps colours as SGR escapes and html draws them as spans in a
<pre>. --trim drops trailing spaces and trailing blank lines.
vorn events --json streams the server's public events, one JSON object per line, until
stopped: {"v":1,"event":"session.exited","at":1760000000000,"session":"…","data":{…}}. The
events are session.created, session.status, session.exited, session.mark,
session.program, agent.hook, git.changed, task.created and task.statusChanged. at is
milliseconds since the epoch and session is present for events about one. --filter a,b keeps
those events; --session <target> keeps one session's. A client that falls behind is coalesced
the way an extension is: only the latest session.status and git.changed of a batch.
A layout is a TOML file in ~/.vorn/layouts/ (the server's data directory), or one an
extension offers with [[contributes.layouts]] and file = "layouts/dev.toml". It is a tree:
a split (split = "horizontal"|"vertical", ratio, [[children]]) or one thing to open: a
session (agent = "claude", command = ["tail", "-f", "app.log"] or shell = true, with
cwd and label) or pane = "browser" with a url. Beside name and project, view is
how the grid shows them: "flexible" (cards laid out as the tree, each split's direction and
ratio as written), "grid" (rows of columns, 1 to 4, automatic without it) or "tabs".
Without view, a tree with splits opens flexible.
name = "dev"
project = "vorn"
view = "flexible"
split = "horizontal"
ratio = 0.6
[[children]]
agent = "claude"
label = "agent"
[[children]]
command = ["tail", "-f", "app.log"]
cwd = "logs"
vorn layout list shows them, vorn layout open <name> [--project X] starts the sessions through
the same calls the app makes and prints their ids (3 when the name is ambiguous), vorn layout save <name> writes the project's open sessions in the view the grid shows them (the flexible grid's
splits and ratios included), and vorn layout check <file> reports every mistake as file:line: message. The app lists layouts under Open Layout › in the command palette. Sessions start in tree
order. In the app, Columns › Flexible in the view menu splits the grid into panes: drag a divider
to resize (double-click evens them), ⌘⌥ arrows move between panes and ⌘⌥⇧ arrows move the divider.
A command runs in a shell, which stays open when it ends; a pane opens beside the session
before it and needs the app to be open.
Workflows run in the server, so a run started here keeps going with nothing open — and a scheduled one happens whether or not the app is running.
Vorn supports two ways of working. Interactive: open a grid of terminals and pair with agents hands-on. Autonomous: define tasks and workflows, and let agents run headlessly while you focus on something else. Mix both freely — they share the same projects, tasks, and git integration.
Run Claude, Copilot, Codex, and Gemini in a responsive grid layout. Resize, reorder, minimize, and filter by status. Focus any terminal fullscreen with one click, or switch to a tab view for sequential browsing. Real-time status detection shows whether each agent is running, waiting, idle, or in an error state.
A session holds its own terminals, a browser, an editor and, with the Device extension from the catalog, an iOS simulator. Draw on the page and the agent sees it; it reads and drives all of them to check its own work.
An agent can publish a page, a doc or a design into its session's browser pane. Select the words you mean and comment on them, edit a doc yourself, and send everything back to the agent as one message. Each publish is a new version, and your earlier comments stay pinned to their words. Workflow review gates use the same viewer, so the steps a gate sends back read your comments one by one. See docs/artifacts.md.
Manage tasks per project with a list view or a drag-and-drop kanban board. Tasks support markdown descriptions, image attachments, git branch targeting, and worktree isolation. Start a task and it launches an agent with the description as the prompt. Tasks can also trigger workflows automatically when created or when their status changes.
Create multi-step workflows with a visual node editor. Chain together agent launches, script executions (Bash, Python, Node.js, PowerShell), and task queue operations. Schedule them manually, once at a specific time, or on a recurring cron. Workflows can also trigger automatically when tasks are created or change status. A workflow can live as a file in its project's .vorn/workflows/ folder, and extensions can add step types and triggers. See docs/workflows.md.
Run agents and scripts in the background without a visible terminal. Headless sessions capture full output logs for later review. Combine with workflow stagger delays to orchestrate multiple agents sequentially, each one picks up where the last left off.
View git changes in a side panel with file-level stats. Click on any changed line to add a review comment, then send all comments to the agent as structured feedback with one click. Stage, commit, and push directly from the terminal session. Full worktree support for safe parallel work on different branches.
Fuzzy search for actions, terminals, recent sessions, projects, and workflows. Quick-launch agents, run workflows, and navigate the app without leaving the keyboard.
Restore previous agent sessions on restart. Vorn tracks session history and matches agent sessions for accurate resumption, so you can pick up right where you left off.
Agents running inside Vorn can access tasks, projects, workflows, sessions, and git state through the built-in MCP server. Create and update tasks, trigger workflows, launch new agents, and query configuration — all from within an agent session.
A minimal always-on-top overlay that stays visible while you work in other apps. See agent status at a glance, answer permission requests and agent questions (including the rules an agent offers to keep), jump back into any session with one click, or fold it to a single line of counts. It is an extension: install the Widget from the catalog in Settings › Extensions; Cmd+Alt+W shows or hides it.
vornd --web); pair a phone or another computer with a code from the desktop appexamples/extensions/turns. An extension runs as a program in any language, or as a sandboxed Wasm component written in Rust with vorn-extension-sdk (vorn ext new --wasm); see docs/extensions.md and docs/wasm-extensions.md. A theme is an extension too: colours for the interface and the terminal, as data; see docs/themes.md| Agent | Command |
|---|---|
| Claude Code | claude |
| GitHub Copilot | gh copilot |
| OpenAI Codex | codex |
| Google Gemini | gemini |
| Hermes | hermes |
Any CLI tool that runs in a terminal works with Vorn. These are the agents Vorn installs by default, with status detection and icons; others, OpenCode among them, install from the catalog in Settings › Extensions.
Prerequisites: a Rust toolchain and Zig 0.16.0 for the native core; Node.js 22 and Yarn for the MCP server, the connector SDK and the tests; brotli for the web client
# Install dependencies
yarn install
# Build vornd, the server, with its session holder and the vorn command
yarn build:core
# Run the desktop app
cd packages/core && cargo run --release -p vorn-app
# Build the web client (the app compiled to WebAssembly) into packages/core/crates/vorn-web/dist
yarn build:web
# Check it in headless Chromium against a test vornd
cd packages/core && node crates/vorn-web/tests/smoke.mjs
Vorn takes contributions as issues. See CONTRIBUTING.md.
MIT - Javier Canizalez