GhostCoder
AI code suggestions in VS Code, powered by your existing Codex account.
Start writing a function, handler, test, type, or comment. When you pause, GhostCoder reads the code around your cursor and suggests a useful multi-line continuation directly in the editor.
No prompt window. No separate AI subscription. No command required in the default mode.
Before you install
GhostCoder requires the Codex CLI to be installed and authenticated.
The Codex VS Code extension by itself is not enough. The codex command must be available on your system, and you must sign in before GhostCoder can generate suggestions.
Follow the official Codex CLI installation guide, then verify the installation:
codex --version
Run Codex once and complete the sign-in flow:
codex
You can confirm that authentication is ready with:
codex login status
If any of these commands fail in your terminal, fix the Codex CLI installation or authentication before troubleshooting GhostCoder.
Install GhostCoder
- Open the Extensions view in VS Code.
- Search for GhostCoder by publisher Vova4o.
- Click Install.
- Run Developer: Reload Window if VS Code asks you to reload.
GhostCoder starts automatically after VS Code reloads.
To install a development build manually, download its .vsix, open the Extensions
view, choose Install from VSIX... from the ... menu, and select the file.
How to use it
Write code normally. GhostCoder waits until typing pauses, then shows a continuation as inline ghost text.
- Accept a suggestion with your normal VS Code inline-suggestion action, usually Tab.
- Keep typing or press Esc to ignore it.
- Click the GhostCoder status-bar item to open diagnostics.
- Run GhostCoder: Pause or Resume from the Command Palette when you need silence.
GhostCoder does not send a request for every character. It debounces typing, cancels obsolete work, suppresses unchanged requests, and limits request frequency.
Choose a model
GhostCoder defaults to gpt-5.3-codex-spark, which is optimized for fast coding responses.
Open VS Code Settings and search for GhostCoder: Model to choose from the included list. You can also run GhostCoder: Select Model to load the models currently available to your Codex account.
The selected model appears in the status bar.
Settings
| Setting |
Default |
Purpose |
ghostcoder.enabled |
true |
Enable or disable GhostCoder. |
ghostcoder.triggerMode |
automatic |
Generate after typing pauses, or only when explicitly requested. |
ghostcoder.model |
gpt-5.3-codex-spark |
Model used for suggestions. |
ghostcoder.reasoningEffort |
low |
Preferred reasoning effort. Unsupported values fall back safely. |
ghostcoder.debounceMs |
750 |
How long typing must pause before a request starts. |
ghostcoder.requestTimeoutMs |
30000 |
Maximum duration of one completion request. |
ghostcoder.maxRequestsPerMinute |
6 |
Rolling request limit. |
ghostcoder.minimumChangeChars |
3 |
Minimum code change before another request is allowed. |
ghostcoder.cooldownAfterAcceptMs |
1500 |
Delay after accepting a suggestion. |
ghostcoder.excludedLanguages |
plaintext, log |
Languages GhostCoder ignores. |
ghostcoder.excludeGlobs |
common secrets, dependencies, generated files |
File patterns GhostCoder ignores. |
ghostcoder.pauseDurationMinutes |
10 |
Duration of a temporary pause. |
There is no bundled keyboard shortcut for generating suggestions. Users who choose manual mode can bind GhostCoder: Generate Code Here themselves.
Diagnostics
Run GhostCoder: Show Diagnostics to see:
- completed, cancelled, suppressed, and failed requests;
- accepted suggestions and accepted characters;
- response latency grouped by model;
- Codex account state and rate-limit information.
Diagnostics remain local and are reset when the extension host restarts.
Privacy and safety
GhostCoder sends a bounded section of code before and after the cursor to the Codex provider configured by your CLI.
To reduce accidental exposure:
- secret-like files such as
.env, private keys, and credential files are blocked;
- languages and file patterns can be excluded in Settings;
- each completion uses an ephemeral Codex thread;
- Codex runs with a read-only sandbox and cannot modify project files;
- tool calls and server-initiated actions are rejected;
- user notification commands and lifecycle hooks are disabled inside GhostCoder by default.
Do not use AI completion on source code that you are not permitted to send to the configured provider.
Troubleshooting
GhostCoder shows no suggestions
- Run
codex --version.
- Run
codex login status.
- Confirm GhostCoder: Enabled is on.
- Confirm the current language or file is not excluded.
- Check whether GhostCoder is paused in the status bar.
- Run GhostCoder: Show Diagnostics.
If the codex executable is not on VS Code's PATH, set GhostCoder: Codex Path to its full path and reload VS Code.
Suggestions appear too often
Increase Debounce Ms, reduce Max Requests Per Minute, or increase Minimum Change Chars.
Suggestions are too slow or weak
Choose a different model or reasoning effort in Settings. Faster models improve latency; stronger models may produce more complete code but take longer.
Development
Requirements: Node.js 20 or newer and VS Code 1.95 or newer.
npm install
npm test
npm run typecheck
npm run build
The opt-in real App Server smoke test consumes Codex usage:
GHOSTCODER_REAL_SMOKE=1 npx vitest run test/realAppServer.smoke.test.ts
