LocalMinds

Privacy-first AI coding assistant for VS Code — offline-capable, multi-model, fully extensible.

LocalMinds gives you a complete AI coding workflow inside VS Code. Run models locally with Ollama for total privacy, connect to 200+ cloud models through a single OpenRouter key, or mix both seamlessly. Your code stays on your machine unless you choose otherwise.

What's New

Project Workflows

The headline feature of this release. Define reusable, multi-step AI tasks that live in .localminds/ alongside your code and are committed with your repo — so your whole team benefits from the same automations.

Building blocks:

Block	Purpose
Instructions	Coding standards and conventions — applied always, by file glob, or on demand
Prompts	Reusable templates with `{{arguments}}` you fill in per run
Agents	A system prompt + model that chats, plans, or edits files
Skills	Focused, on-demand expertise an agent can invoke

Chaining and loops: Assemble blocks into a Workflow — ordered steps where each can reference an earlier step's output via {{steps.ID.output}}. Repeat steps a fixed number of times or loop until a condition is met.

Run experience: Live streaming output per step, a diff preview you approve before any file is changed (or enable auto-apply), and a Cancel button to stop at any point.

Run LocalMinds: Project Workflows from the Command Palette or the chat title bar. On first run, a small example workflow is seeded in .localminds/ to get you started.

Plan Mode

A new dedicated chat mode built for structured thinking before writing code.

Produces a concrete, step-by-step implementation plan you can review and revise.
Ideal for architecture decisions, refactors, and scoping larger features.
No file edits are applied — Plan mode is intentionally read-only, keeping exploration separate from execution.

Switch to it from the mode dropdown in the chat panel header (see Chat modes below).

Context Window Indicator

A ring gauge beside the input box shows how much of the model's context window is in use, broken down into three layers:

Segment	What it represents
History	Previous messages in the current session
System prompt	Agent profile, instructions, and mode prompt
Files	Active file, pinned files, and gathered workspace context

Hover over the ring for a detailed tooltip with exact token counts per segment. Use it to decide when to start a new chat or to diagnose unexpectedly slow responses.

Per-Call Model Override and Independent Abort

Workflow steps can now override the active model on a per-step basis — useful for running a cheap fast model for boilerplate steps and a more capable model only where it matters.

Each step can also be cancelled independently. Aborting a workflow step does not interrupt a concurrent chat panel stream running in parallel.

Mode Selector Redesigned

The previous toggle buttons have been replaced with a dropdown menu that shows a short description of each mode alongside its name — so the difference between Ask, Plan, and Agent is always visible without needing to remember.

"Chat" Renamed to "Ask"

The general-purpose conversational mode is now called Ask for clearer naming alongside Plan and Agent modes.

Why LocalMinds?


Privacy first	Local models keep your code on your machine. No telemetry, no tracking, no data collection.
Cost effective	Run Gemma 4 locally for free, or use `openrouter/free` for zero-cost cloud models. Pay-as-you-go when you need more power.
Zero latency	Local inference means instant responses with no network round-trips.
One key, every model	A single OpenRouter API key unlocks Claude, GPT, Gemini, Llama, Grok, and hundreds more. One bill, no juggling provider accounts.
Extensible workflows	Define reusable prompts, agents, and multi-step workflows that live in your repo and run with a click.

Features

Chat modes

LocalMinds provides three distinct modes, each optimised for a different stage of development:

Ask — conversational Q&A, explanations, and code walkthroughs.
Plan — structured implementation plans, architecture decisions, and step-by-step breakdowns before you write a line of code.
Agent — autonomous file edits with a diff preview you approve before anything is applied.

Switch modes from the dropdown in the chat panel header.

Project Workflows

Build repeatable, multi-step AI tasks that are committed with your repo and shared with your team.

Building blocks:

Instructions — coding standards and conventions applied always, by file glob, or on demand.
Prompts — reusable templates with {{arguments}} you fill in per run.
Agents — a system prompt + model that chats, plans, or edits files.
Skills — focused, on-demand expertise an agent can invoke.

Chaining: Assemble blocks into a Workflow — ordered steps where each can reference an earlier step's output via {{steps.ID.output}}, plus loops that repeat a fixed number of times or until a condition is met.

Run experience: Live streaming output, a diff preview you approve before any file is changed (or enable auto-apply), and a Cancel button to stop at any point.

Run LocalMinds: Project Workflows from the Command Palette or the chat title bar. First run seeds a small example to get you started. Everything is stored in .localminds/ at your project root.

Context window indicator

A ring gauge beside the input box shows how much of the model's context window you are using, broken down by chat history, system prompt, and attached files. Hover for a detailed tooltip.

Agent Profile

Run LocalMinds: Customise Agent Profile to describe your role, stack, coding style, and preferences. The profile is appended to every system prompt so responses fit the way you work.

Smart context

LocalMinds automatically gathers the active file, selection, and related workspace files. Pin additional files with the attach button for precise control.

Editor commands

Right-click any selection to access Generate, Refactor, Explain, or Inline Edit (custom instruction).

Keyboard shortcuts

Shortcut	Action
`Cmd+Shift+G`	Generate code from a description
`Cmd+Shift+E`	Inline edit the current selection

Offline mode

Click the status-bar item or set localminds.offlineMode: true — all cloud calls are skipped and only local models are used.

Chat history

Review, search, and resume previous conversations from the history panel.

Getting started

Prefer a guided flow? Run LocalMinds: Open Setup Wizard from the Command Palette (Cmd+Shift+P) after installing.

1. Install Ollama

# macOS
brew install ollama          # or download from https://ollama.com

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows — download from https://ollama.com/download

2. Pull a local model

Model	Size	VRAM	Best for
`gemma4:e2b`	7.2 GB	~6 GB	Laptops with limited memory
`gemma4:e4b`	9.6 GB	~8 GB	Good balance (default)
`gemma4:12b`	7.6 GB	~8 GB	256K context on modest RAM
`qwen3-coder:30b`	19 GB	~24 GB	Best for coding
`gemma4:26b`	18 GB	~20 GB	Strong general-purpose
`gemma4:31b`	20 GB	~24 GB	Maximum Gemma capability

ollama pull gemma4:e4b
# or for coding on a 24 GB+ machine:
ollama pull qwen3-coder:30b

3. (Optional) Add an OpenRouter key for cloud models

One key unlocks Claude, GPT, Gemini, and hundreds more — including free models.

Sign up at openrouter.ai
Create a key at openrouter.ai/keys
Optionally add credit for premium models, or use openrouter/free at no cost

Skip this step entirely for offline-only operation.

Configuration

Install the LocalMinds extension, then open Settings and search LocalMinds — or add directly to settings.json:

{
  "localminds.ollama.model": "gemma4:e4b",
  "localminds.openrouter.apiKey": "sk-or-v1-...",
  "localminds.openrouter.model": "openrouter/free",
  "localminds.openrouter.models": [
    "openrouter/free",
    "google/gemma-4-31b-it:free",
    "anthropic/claude-sonnet-4.6",
    "openai/gpt-4o",
    "google/gemini-3.5-flash"
  ]
}

Popular cloud models

See the full catalogue at openrouter.ai/models.

Use case	Model ID
Free cloud models	`openrouter/free`
Free cloud (Gemma)	`google/gemma-4-31b-it:free`
Best all-round coder	`anthropic/claude-sonnet-4.6`
Heavy reasoning	`anthropic/claude-opus-4`
Cheap and fast	`openai/gpt-4o-mini`
Long context	`google/gemini-3.5-flash`
Open-weights	`meta-llama/llama-3.1-70b-instruct`

Recommended hardware

Setup	Model	Experience
MacBook Air M1 (8 GB)	`gemma4:e2b`	Usable, ~5–10 s
MacBook Pro (16 GB)	`gemma4:e4b`	Good, ~3–5 s
MacBook Pro (32 GB+)	`qwen3-coder:30b`	Best local coding
GPU with 24 GB+ VRAM	`qwen3-coder:30b` / `gemma4:31b`	Best local experience

Cloud models via OpenRouter have no local hardware requirements.

Troubleshooting

Ollama not connecting — make sure the server is running with ollama serve, then verify with curl http://localhost:11434/api/tags.

Model not found — run ollama list to see installed models, then ollama pull gemma4:e4b to download.

OpenRouter errors:

401 — verify your API key in localminds.openrouter.apiKey matches a key at openrouter.ai/keys.
402 — add credit at openrouter.ai/credits.
429 — rate limited; wait a moment or switch to a different model.
Model not found — IDs must match exactly (e.g. anthropic/claude-sonnet-4, not claude-sonnet-4).

Slow responses — try a smaller Gemma variant, reduce localminds.contextLines to 50, or switch to a cloud model.

LocalMinds

LocalMinds — Local & Cloud AI Coding Assistant

LocalMinds

What's New

Project Workflows

Plan Mode

Context Window Indicator

Per-Call Model Override and Independent Abort

Mode Selector Redesigned

"Chat" Renamed to "Ask"

Why LocalMinds?

Features

Chat modes

Project Workflows

Context window indicator

Agent Profile

Smart context

Editor commands

Keyboard shortcuts

Offline mode

Chat history

Getting started

1. Install Ollama

2. Pull a local model

3. (Optional) Add an OpenRouter key for cloud models

Configuration

Popular cloud models

Recommended hardware

Troubleshooting

Changelog

License

localminds