Requires VS Code 1.120+ with GitHub Copilot Chat. Install from the Marketplace (kelviny.lai-compatible-copilot) or build a VSIX with npm run build. API keys live in VS Code Secret Storage, never in settings.json.

What it does

Copilot Chat normally talks only to GitHub's own models. This extension registers its own model provider, so the same chat UI can drive:

• Local models — llama.cpp, Ollama, vLLM
• Cloud models — OpenAI, Anthropic, Gemini, DeepSeek, Qwen, and any OpenAI-compatible endpoint

Quick start

1. Zero config: the built-in free models

On a fresh install the picker already lists a set of free models shown as 🌙 cline … (the moon distinguishes them from your own ⭐ cloud models). They use a shared key bundled with the extension — nothing to configure. See Built-in free models.

2. Run a local server

# llama.cpp — --jinja lets the server render tool-call templates,
# --mmproj adds vision, --ctx-size must match your laicopilot.models[].context_length
llama-server --model /path/qwen3.6.gguf --mmproj /path/mmproj-f16.gguf \
  --host 0.0.0.0 --port 8080 --n-gpu-layers 999 --ctx-size 131072 --jinja

ollama serve                                    # Ollama
vllm serve Qwen/Qwen3-35B-A3B --port 11000     # vLLM

3. Add the model

In settings.json (or the configuration UI):

{
  "laicopilot.models": [
    {
      "id": "qwen3.6",
      "owned_by": "llamacpp",
      "apiMode": "llamacpp",
      "displayName": "Qwen 3.6 (local)",
      "baseUrl": "http://localhost:8080",
      "vision": true,
      "context_length": 131072,
      "max_tokens": 65536,
      "thinking_budget": 1024,
      "max_history_messages": 20
    }
  ]
}

Pick the model in the Copilot Chat model dropdown and start chatting — tools are called automatically.

Cloud models

Declare an endpoint once in laicopilot.cloudProviders and the extension fetches its model list automatically; fetched models are tagged ⭐ <provider> <model> and route to that endpoint and its own key:

{
  "laicopilot.cloudProviders": [
    { "id": "deepseek", "baseUrl": "https://api.deepseek.com", "apiMode": "openai" },
    { "id": "cline", "baseUrl": "https://api.cline.bot/api/v1", "modelFilterKeywords": ["free"] },
    { "id": "anthropic", "baseUrl": "https://api.anthropic.com", "apiMode": "anthropic" }
  ]
}

Prefer explicit entries? Put them in laicopilot.models instead:

{
  "laicopilot.models": [
    { "id": "gpt-4o", "owned_by": "openai", "baseUrl": "https://api.openai.com/v1", "apiMode": "openai" }
  ]
}

baseUrl may include or omit /v1 — the extension never doubles it (/v1/v1/…).

Built-in free models

A preset ships with the extension so a fresh install works out of the box:

Notes:

• These models run on the upstream free tier, so availability and rate limits are decided upstream. If one errors or is throttled, switch to another free model.
• The shared key is public by design: it is embedded in the published extension. To use your own quota, store a key for cline (it always wins over the bundled one).
• If you configure your own provider with id cline, your configuration wins; the bundled key is only reused while it still points at the official base URL, so it is never sent to an unknown host.
• Set laicopilot.enableDefaultFreeProvider to false to hide them.

API keys

The model list never prompts for a key. A provider without a key simply appears as a ⚠ <provider> needs API key row in the picker. A key is requested only when you actually send a chat request to that provider (or when you run one of the commands below).

Leaving the input box empty and pressing Enter clears the stored key.

Storage in Secret Storage (OS keychain / Windows Credential Manager):

If a stored key for the same provider exists, it always takes priority over the bundled free key. Local servers (llama.cpp / Ollama / vLLM) need no key; if yours runs with --api-key sk-xxx, store it under the provider name.

When the upstream rejects a key (401/403) or rate-limits it (429), the extension shows a notification with Add my API key / Open configuration / Hide built-in free models buttons and flags the status bar, instead of dumping the raw upstream error into the transcript.

Settings

Global

Per model (laicopilot.models[])

Identity & routing

Context & tokens

On a context-window error the extension automatically retries with fewer messages and truncated tool results — no setting needed. To compact long conversations proactively, enable laicopilot.autoCompact (global setting, see above).

Sampling & reasoning

Configuration UI

laicopilot.openConfig opens a Webview panel (no JSON editing required):

• 📋 Add / edit / delete models — validated form for every field, no JSON editing
• ☁️ Manage cloud providers — add, edit and delete provider endpoints together with their keys
• 🔑 Key management — type keys straight into Secret Storage
• 🔎 Fetch models — pull the live model list of a provider and pick from it instead of typing an ID
• ⚙️ Global settings — delay, retry, auto-compaction, log level, commit language, read_file line count
• ⇄ Export / import — move your whole setup (models, providers, keys) between machines; note the exported JSON holds keys in plain text

Git commit messages

Mark one model with "useForCommitGeneration": true (or pick it in the configuration UI), then click the wand button in the Source Control toolbar — laicopilot.generateGitCommitMessage does the same from the command palette. laicopilot.commitLanguage and laicopilot.commitMessagePrompt control the output; laicopilot.abortGitCommitMessage stops generation.

Status bar & logging

The L-AI status bar item shows the token usage of the current request (click it to open the configuration UI) and turns red when a key is missing or rejected. Logs are JSON lines under ~/.copilot/laicopilotdev/logs/; enable them with laicopilot.logLevel: "debug".

Troubleshooting

License

MIT — see LICENSE. Release notes: docs/CHANGELOG.md.