Wilbur for VS Code
Wilbur — a local coding agent for the terminal, tools, subagents, a plan
it keeps, driven entirely by models on your own GPU. This extension runs
Wilbur in a VS Code terminal panel, the way Claude Code's own VS Code
integration works, without sending anything to a cloud model.
GitHub: https://github.com/jsonholdings/wilbur
Requirements
- Ollama running locally, with a tool-capable model pulled.
- The
wilbur CLI installed and on your PATH (see the main repo's README
for the installer). If VS Code can't find it, set wilbur.path.
Install
From the Marketplace: search Wilbur and click Install, or from Open VSX
if you're on a Marketplace-compatible fork (VSCodium, etc).
From a .vsix (built from source):
cd vscode
npx @vscode/vsce package
code --install-extension wilbur-vscode-0.3.0.vsix
(On a Flatpak VS Code install, prefix both commands with
flatpak-spawn --host.) Reload the window afterwards
(Developer: Reload Window).
Use
- Wilbur: Open (command palette, or the Wilbur icon in the activity bar) —
opens a terminal running
wilbur with its cwd set to the workspace root.
- Wilbur: Ask about selection (right-click a selection, or
ctrl+alt+w) —
sends the file path, line range and selected text into the Wilbur terminal
as a prompt, starting Wilbur first if it isn't already running.
- Wilbur: Send current file — sends the active file's path into the
Wilbur terminal.
- Wilbur: Resume last session — launches
wilbur --continue in a new
terminal, resuming the latest saved session for the workspace's cwd.
- Wilbur: Sessions… — lists wilbur's saved sessions (via
wilbur --sessions --json) in a quick pick, showing each one's objective,
working directory and how long ago it was updated, then resumes the one
you pick with wilbur --resume <id> in a new terminal.
- Wilbur: Send problems in file — sends every diagnostic (error/warning)
in the active file to Wilbur as one message.
- Fix with Wilbur — a quick-fix code action offered on any diagnostic
(error/warning) that sends its file, line and message to Wilbur.
- Status bar model indicator — shows the model wilbur will launch with
and whether Ollama is reachable (polled every 30s). Click it to list local
Ollama models and switch wilbur's default model
(writes
~/.config/wilbur/config.json directly; applies to the next
launch, not a running terminal).
Settings
| Setting |
Default |
Meaning |
wilbur.path |
wilbur |
Path to, or name on PATH of, the wilbur executable, resolved on the host. If it can't be found, Wilbur shows an error with a link to Settings. |
wilbur.useFlatpakSpawn |
auto |
auto prefixes flatpak-spawn --host only when VS Code itself is running in a Flatpak sandbox (detected via /.flatpak-info); on/off force it. |
wilbur.model |
"" |
Passed to wilbur via -m on launch when set. Empty uses wilbur's own default. The status bar's model switcher edits wilbur's config file directly instead of this setting. |
wilbur.approvalMode |
manual |
auto passes --yes to wilbur, auto-approving its proposed actions. |
wilbur.contextSize |
0 |
Passed to wilbur via -c when greater than 0. 0 means unset. |
Why flatpak-spawn --host
If your VS Code install is a Flatpak (e.g. com.visualstudio.code), the
extension host runs inside that sandbox. wilbur (and node/npm used to
build this extension) live on the host, not in the sandbox, so launching
wilbur directly from inside VS Code would fail with "command not found".
flatpak-spawn --host <cmd> runs <cmd> in the real host namespace instead.
This is auto-detected and only applies on Flatpak installs.
Development
npm test
Unit tests cover the sandbox-detection/launch-argv logic in src/sandbox.js
without requiring the vscode module.
Status
v1: terminal-based (open, ask-about-selection, send-current-file). Diff
accept/reject for Wilbur's proposed edits is planned — see BACKLOG.md.
License
Apache-2.0. See LICENSE.
