Gemma Agent

Local AI coding assistant for VS Code — chat, inline completion, in-place editing, agentic file editing, semantic workspace search and commit message generation. 100% offline: your code never leaves your machine.

Runs on any local model server: Ollama (recommended) or any OpenAI-compatible server — LM Studio, Jan, llama.cpp llama-server, vLLM, LocalAI. Lives in its own sidebar, like Copilot/Cursor.

Features

🚀 Guided setup

• A setup wizard detects whether your server is installed, running, and has a model — and walks you through each step (install, start, download a model with a progress bar). No manual fiddling.
• Provider cards switch between Ollama and OpenAI-compatible servers; the UI adapts to each backend's capabilities.

💬 Chat

• Streaming chat in the sidebar with syntax highlighting, copy / insert-into-editor buttons
• Multiple sessions — name, switch, rename and delete conversations; everything (including tool cards) persists per workspace
• Slash commands — type / for /explain, /fix, /tests, /docs, /clear; your selection or active file is attached automatically
• #file / #terminal references — type # to attach a workspace file or the last terminal command's output
• @workspace — semantic search over a local embedding index of your codebase (opt-in)
• Automatic context-window management: long conversations are trimmed to fit the model instead of silently breaking it
• Custom instructions via a setting or a .gemma/rules.md file

⚡ Agent mode

Let the model create files, edit code and run commands in a tool-call loop:

• Approval flow — shell commands (and optionally file writes) wait for your Approve / Deny / Always-allow; View diff previews a change before you approve it
• Review changes — after a run, review every changed file and Keep or Revert each, or Undo all at once
• Auto-verify — after edits the agent runs diagnostics and fixes its own errors before finishing
• Live tool cards with status, collapsible output, and a step indicator

✏️ Inline edit

Select code (or just place the cursor) and press Cmd+Shift+I / Ctrl+Shift+I:

• The rewrite streams directly into your editor, highlighted as it arrives
• Accept with Cmd+Enter / reject with Esc (CodeLens buttons too)
• The whole edit is a single undo step

👻 Inline completion

• Ghost-text completions with debouncing and careful post-processing
• Accept word-by-word (Cmd+→ / Ctrl+→), whole suggestion (Tab), or cycle alternatives (Alt+] / Alt+[)
• Per-language enable/disable (gemmaAgent.completionLanguages)

🛠 Extras

• Commit messages — ✨ button in the Source Control title bar writes a conventional-commit message from your staged diff
• "✨ Fix with Gemma" quick-fix on any error squiggle (no selection needed)
• Right-click code actions: Explain, Refactor, Fix, Generate Tests
• Terminal helpers: explain/fix the last command (captured automatically via shell integration), or run a selection
• Commit messages — ✨ button in the Source Control title bar from your staged diff
• Server lifecycle from the status bar and sidebar (start, stop, pull models with progress, model picker)

Requirements

• A local model server. Easiest: Ollama — the extension can install-guide and pull models for you.
• Or any OpenAI-compatible server (LM Studio, Jan, llama.cpp llama-server, vLLM, LocalAI): set gemmaAgent.apiProtocol to openai-compatible and point gemmaAgent.ollamaUrl at it.
• A model, e.g. gemma4:e4b. For @workspace, also pull an embedding model (nomic-embed-text).

Choosing models (this matters as much as the code)

Gemma Agent routes each task to a role, so you can run a different model per job. Set these in Settings; empty roles fall back to the main model.

Role	Setting	Good picks (mid‑2026)	Why
Chat	gemmaAgent.model	gemma4:e4b, qwen3:8b, gemma3:12b	General Q&A, explanations
Agent	gemmaAgent.agentModel	qwen3-coder:30b, devstral:24b, qwen2.5-coder:7b	Reliable tool‑calling + multi‑file edits
Completion	gemmaAgent.completionModel	qwen2.5-coder:1.5b-base, :3b-base, codegemma:2b	Real FIM ghost text — small and high‑quality
Embedding	gemmaAgent.embeddingModel	nomic-embed-text, embeddinggemma	@workspace semantic search

Notes:

• Ghost text quality comes from a code model with fill‑in‑the‑middle, not from size — qwen2.5-coder:1.5b-base is tiny and excellent. Use the -base variant for FIM. Detection is automatic (gemmaAgent.completionFim: auto).
• Agent reliability scales with the model. A strong tool‑caller plus structured‑output (gemmaAgent.agentStructuredOutput, on by default where supported) makes agent mode far more dependable than a small general model.
• Everything installs the same way: ollama pull qwen2.5-coder:1.5b-base. If a role's model isn't installed, Gemma Agent offers to pull it and falls back to the main model meanwhile.
• The starting recommendations live in src/llm/modelCatalog.ts — keep it current; local models move fast (e.g. Qwen3‑Coder‑Next, DeepSeek V4, MiniMax M3, Kimi K2, Gemma 4 12B). You can run any model regardless of the catalog.
• Faster generation: for a big speedup, run a server that supports speculative decoding with a small draft model — LM Studio (≥0.3.10) or llama-server --model-draft. Ollama doesn't expose this yet. gemmaAgent.keepAlive (default 30m) keeps the model warm between turns.

Keyboard shortcuts

Action	macOS	Windows/Linux
Open chat	⌘⇧G	Ctrl+Shift+G
Explain selection	⌘⌥E	Ctrl+Alt+E
Inline edit (selection or cursor)	⌘⇧I	Ctrl+Shift+I
Accept inline edit	⌘⏎	Ctrl+Enter
Reject inline edit	Esc	Esc
Accept completion	Tab	Tab
Accept next word of completion	⌘→	Ctrl+→
Cycle completion alternatives	⌥] / ⌥[	Alt+] / Alt+[
Stop generation (chat)	Esc	Esc

Key settings

Setting	Default	Description
gemmaAgent.apiProtocol	ollama	ollama or openai-compatible
gemmaAgent.model	gemma4:e4b	Main chat model
gemmaAgent.agentModel	""	Model for ⚡ Agent mode; empty = main model
gemmaAgent.completionModel	""	Model for inline completion; empty = main model
gemmaAgent.agentStructuredOutput	true	Schema-constrain agent tool calls for reliability
gemmaAgent.completionFim	auto	Real fill-in-the-middle when the model supports it
gemmaAgent.keepAlive	30m	Keep the model warm in RAM between requests
gemmaAgent.numCtx	8192	Context window size (more = more RAM)
gemmaAgent.agentRequireApproval	commands	Which agent tools need approval (commands / commandsAndWrites / never)
gemmaAgent.agentAutoVerify	true	Run diagnostics + self-fix after agent edits
gemmaAgent.workspaceIndexEnabled	false	Enable @workspace semantic search
gemmaAgent.autoWorkspaceContext	false	Auto-attach relevant code to chat (needs index)

See the full list in Settings under Gemma Agent.

Privacy

Everything runs locally through your own model server. No telemetry, no cloud calls, zero runtime dependencies. (OpenAI-compatible servers are assumed keyless/local; no API key is sent.)

License

MIT