01
The Agents Window
A collapsible sidebar, a centred chat column, and a tool rail that folds to a 40px strip. The branch and your remote runners sit at the composer.
v0.2.0MITWindows · macOS · Linux
Vellum Code — an agent that shows its work
Vellum Code reads your project, makes the change, verifies it, and reports back — with the whole operation visible. GUI‑first and local‑first, MCP‑native, seven providers, and a CLI that shares the same brain.
Signed installer · auto-updates in-appNo VSCode requiredYour keys stay on your machine
What it does
Everything visible.
A chat-first agent that reads your project, makes the change, verifies it, and reports back — with the whole operation on screen. No black boxes, no silent edits.
02
Changes panel
The working tree as a reviewable surface: staged and unstaged groups, per-hunk stage, unstage and discard, and a deterministic branch-and-commit.
03
Parallel sessions
Switching chats never kills a running turn. It keeps streaming behind an amber dot, and returning catches you up mid-sentence.
04
The agent loop, in the open
Streaming tool calls, and a live-writing view where the file grows inside the card as the agent types. Auto-compact before you hit the wall, a files card with undo and review, stuck detection, a wall-clock watchdog, and provider failover.
05
Live Preview
A real browser pane in the rail. Dev-server URLs are sniffed from terminal output, and a runtime error becomes a one-click fix request.
06
Trusted apply
Edits that match through whitespace and line-ending drift and refuse when a pattern hits twice, edit-time diagnostics from a language server, and reject-with-instruction: say how it should have been done and the retry obeys.
07
/undo and /redo
Snapshot rollback for the whole workspace. Restoring invalidates the agent's read-state, so it never works from a stale picture.
FAQ
The questions that matter.
Do I need a subscription?
No. The app is MIT-licensed and free. You pay your model provider directly, with your own key — there is no account here to buy.
Does it need VSCode?
No. It is a standalone desktop application. It can load VSCode themes and grammars, but there is no extension host and nothing to install into an editor you already have.
Is my code sent anywhere?
Only to the model provider you configured, and only what the turn needs. Keys are stored encrypted on your machine, an audit log records every outbound request as metadata, and a local-only mode blocks egress that is not on your allow-list.
Can I run local models?
Yes. Anything that speaks the OpenAI-compatible API works, including a local Ollama daemon. The semantic index has a keyless fallback, so search works without a provider at all.
How does it handle two tasks at once?
Each chat keeps its own agent. Switching does not interrupt a running turn — it goes on streaming behind an amber dot, and /tell steers it without switching back. Longer work can be dispatched to the fleet, which runs in an isolated git worktree.
What about very large repositories?
Reads return a window with the file’s symbol outline rather than the whole file, the context compacts itself before it fills, and repo-wide questions are delegated to a child agent so the file dumps land in its context instead of yours.
Get started
Point it at a repo.
Install the desktop app, run vellum init, and give it a first task. It will show you everything it does next.
Or the CLI, same engine
npm i -g vellum-codevellum init && vellum