LLMcode
Run opencode inside VS Code — as the model behind Copilot's chat, as a @llmcode agent, and as a native code assistant.
LLMcode does not ship its own chat UI. It plugs the opencode engine into VS Code's existing surfaces so everything runs on your opencode models, providers, agents, skills and MCP servers — with a 1M-token context window.
What it does
LLMcode integrates with VS Code in three ways:
| Integration |
How it shows up |
What it powers |
Language-model provider (vendor: llmcode) |
Pick LLMcode (opencode) in the model picker of the native chat |
Copilot Chat, inline chat, terminal chat, and the Agents window all run on opencode models |
Chat participant (@llmcode) |
Type @llmcode … in chat |
An opencode-backed agent inside the chat, alongside the built-in participants |
| Code assistant |
Ghost-text completions, Ctrl+K inline edit, editor context menu |
Tab autocomplete, inline edits of a selection, and /explain, /fix, /refactor, /test, /doc commands |
All configuration — providers, models, agents, skills, MCP servers, plugins, general options — lives in one LLMcode Settings panel. There is no separate chat sidebar to learn: you work in VS Code's own chat.
The extension uses only native VS Code APIs: vscode.l10n for localization, VS Code color themes, ThemeIcon, QuickPick, createTerminal, vscode.diff, the lm language-model API, and chat.createChatParticipant.
How it uses opencode
LLMcode talks to a managed opencode serve process through the official @opencode-ai/sdk:
- On activation it finds the
opencode CLI, starts the engine, and (if the CLI is missing) offers to install it.
- It starts on the first free port (falling back from
4096 if a stale server occupies it), so it never fails with ServeError.
- Context and output limits are read from the engine, so a model that reports a 1,000,000-token context window is exposed to VS Code as such.
- Provider secrets stay in your opencode configuration — LLMcode never copies them.
Requirements
- VS Code
^1.94.0 (stable APIs only — no proposed API needed)
- The
opencode CLI. If it is not installed, run LLMcode: Check opencode Installation and the extension will offer to install it through the integrated terminal.
Quick start
- Install the extension (Marketplace:
llmcode.llmcode, or code --install-extension llmcode.llmcode).
- Reload the window so the extension activates on startup.
- Open the chat and either:
- open the model picker and choose a model under LLMcode (opencode), or
- type
@llmcode followed by your question.
- Ask anything — the response is streamed from the opencode engine.
See docs/USAGE.md for the full guide (configuration, all commands, settings, troubleshooting, verification).
Verifying that LLMcode (not Copilot) is answering
- Open the chat and its model picker.
- Pick any model under LLMcode (opencode).
- Ask a question — the response comes from opencode.
Programmatically: vscode.lm.selectChatModels({ vendor: 'llmcode' }) returns the opencode model list. This is asserted in the test suite (streams a response from opencode through the provider).
Note on "replacing Copilot": VS Code's Agents window is driven by Copilot's own internal harnesses, which a third-party extension cannot replace. What LLMcode can do — and does — is register opencode as a model provider and as a @llmcode participant, and set chat.defaultModel to an opencode model so the native chat uses it by default. Once you pick an LLMcode model, the chat runs on opencode.
Commands
| Command |
Default key |
Description |
LLMcode: Open Chat |
Ctrl+Shift+L |
Opens VS Code's native chat |
LLMcode: Select Chat Model |
— |
Opens the native model picker to choose an LLMcode model |
LLMcode: Show Model Log |
— |
Opens the LLMcode output log (which model actually answered each completion) |
LLMcode: Use Local Agent by Default |
— |
Makes the Agents window default to Local instead of Copilot |
LLMcode: Open Settings |
— |
Opens the LLMcode Settings panel |
LLMcode: New Session |
— |
Opens the chat to start a new session |
LLMcode: Import Skills/MCP/Plugins from Machine |
— |
Imports from ~/.config/opencode, project .opencode, and WSL |
LLMcode: Inline Edit Selection |
Ctrl+K |
Rewrites the selected code via opencode |
LLMcode: Explain / Fix / Refactor Selection |
— |
Selection commands (editor context menu) |
LLMcode: Toggle Autocomplete |
— |
Enables/disables ghost-text completions |
LLMcode: Take Over Inline Suggestions |
— |
Disables Copilot autocomplete and enables LLMcode's |
LLMcode: Check opencode Installation |
— |
Detects the CLI, offers to install |
LLMcode: Install opencode |
— |
Runs the installer in the integrated terminal |
LLMcode: Open Terminal Session |
— |
Opens a terminal running opencode |
Settings
| Setting |
Default |
Description |
llmcode.autocomplete.enabled |
true |
Enable tab autocomplete (ghost text) |
llmcode.autocomplete.model |
"" |
Model for autocomplete, e.g. abliteration/opus-5.5. Empty uses the default |
llmcode.mode |
"auto" |
chat, assistant, or auto |
llmcode.opencode.command |
"opencode" |
Path or name of the opencode executable |
llmcode.opencode.port |
0 |
Fixed port for opencode serve (0 = pick a free port) |
llmcode.opencode.hostname |
"127.0.0.1" |
Hostname for opencode serve |
llmcode.setDefaultChatModel |
true |
Set chat.defaultModel to an LLMcode model when none is configured |
Development
npm install
npm run compile # esbuild bundle
npm run watch # rebuild on change
npm run typecheck # tsc --noEmit
npm test # extension-host test suite (13 tests)
npm run package # vsce package (.vsix)
Press F5 to launch an Extension Development Host.
License
MIT