Overview

Northstar is a multi-agent coding assistant that lives inside VS Code. Unlike single-turn AI assistants, Northstar can plan, act, and self-correct — executing sequences of real tool calls (file edits, terminal commands, browser searches, git operations) until a goal is complete.

It works with any OpenAI-compatible endpoint: local models via Ollama or llama.cpp, OpenRouter's free-tier models, or paid providers like OpenAI and Anthropic. No mandatory subscription. No data sent anywhere you haven't configured.

What's new in 1.6

• Tool approval ladder — choose PLAN / MANUAL / AUTO EDIT / FULL AUTO from the chat footer instead of answering a popup for every tool call.
• Plans you approve — in PLAN mode the agent researches, writes its plan to .northstar/plans/, and waits for Proceed before changing anything.
• See the change before you approve it — approval cards show the diff inline; Diff opens the whole change in VS Code's diff editor.
• Autopilot — give Northstar a list of tasks and let it work through them while you're away, each on its own git branch for you to review.
• A sandbox that actually runs — terminal commands are sandboxed by default: writes stay inside your workspace and credential stores are unreadable.
• Safer parallel agents — each parallel agent works in its own git worktree, so agents can't overwrite each other's edits.
• More reliable browser — a visible browser window (falls back to your installed Chrome), plus browser_wait_for and browser_network_log.
• AGENTS.md support — project instructions are read from NORTHSTAR.md, AGENTS.md or CLAUDE.md.
• The full agent prompt now ships with the extension — earlier releases ran on a one-line fallback prompt.

See the Changelog tab on the Marketplace page for the full list of fixes.

Features

Agent Mode

Event driven Model and Agent Driven Model for the execution, User can choose one.

Multi-Agent Task Execution

Spawn multiple AI sub-agents in parallel, each with its own goal and toolset. Each parallel agent runs in its own git worktree on its own branch, so agents can't overwrite each other's edits; their work is left on branches for you to review.

Background Agent

Run multiple tasks in the background using background agents. Task status is displayed in the Northstar panel.

Tool Approval Ladder

Pick how much Northstar may do without asking from the selector in the chat footer:

PLAN is enforced at the tool layer — sub-agents included — not just requested in the prompt. Tools from external MCP servers ask for approval unless you're on FULL AUTO or have granted Full Access in the permission settings.

Plan Mode with Reviewable Plans

In PLAN, the agent researches the task and saves a plan to .northstar/plans/<name>.md. The plan appears in the chat with Proceed and Discard. Proceed records your approval, switches to AUTO EDIT and hands the agent the plan file, so it builds exactly what you approved.

Diff Preview Before Approval

Approval cards for file edits show the change itself — lines added and removed, plus the first lines of the diff. Diff opens the full change side by side in VS Code's diff editor. The preview is computed the same way the edit will be applied, so what you approve is what gets written.

Autopilot

Open the ★ Autopilot panel, add tasks, and press Start. Northstar works through the approved list on its own:

• Each task runs in a separate git worktree on its own branch — your working tree is never touched and nothing is pushed.
• Unattended, the agent may only read, edit, run tests/builds and commit locally. Anything else (browser, network, installs, pushes, deleting outside the worktree) parks the task with a note explaining why.
• Each finished task shows its branch, lines changed and test result. Review opens the branch diff; Merge and Discard ask before acting.
• With nothing approved, Northstar reads the codebase and adds suggestions to a Proposed list — it never builds them until you approve one.
• Limits (spend per run and per task, tasks per run, minutes per task) and an optional nightly window live under Settings → Northstar → Autopilot.

Autopilot runs only while VS Code is open and the machine is awake. Providers that report no cost (such as Ollama) make the spending limits inactive — the minutes-per-task limit still applies.

Universal LLM Support

Connect to any provider with a single URL and API key:

• Local — Ollama, llama.cpp, LM Studio, or any OpenAI-compatible server
• Cloud — OpenAI, Anthropic, OpenRouter, and Ollama Cloud (including free-tier models like gpt-oss:120b-cloud)
• Automatic model detection — Northstar probes your endpoint and selects the best available model

Full Tool Suite

Northstar agents act on your codebase — not just talk about it:

Workstream Live View

Watch every tool call in real time as agents execute. See what's being read, written, searched, or run — with a full expandable activity log per session.

RAG — Workspace Intelligence

Northstar indexes your project with LanceDB (local vector database) and BM25 for hybrid retrieval. Agents automatically pull relevant context from your codebase before responding — no manual file pasting required.

Persistent Conversation Memory

Conversations are saved locally and restored on re-open. Switch between past sessions without losing context. Northstar links session history to the workspace it was opened in.

Sandbox

Terminal commands run inside an OS sandbox (macOS sandbox-exec, Linux unshare), on by default. Writes are confined to your workspace and temp folders, and credential stores (~/.ssh, ~/.aws, ~/.gnupg, keychains and others) can't be read. Turn it off in the Northstar settings, or per workspace with .northstar/sandbox.json.

Thinking Mode

Toggle deep reasoning for complex architectural problems, refactoring decisions, or multi-step debugging. The agent narrates its reasoning before acting, making it easier to course-correct.

Auto Compaction

Long sessions are compacted automatically so you can keep working in the same conversation.

Skills

Northstar supports imported skills and built-in ones such as Todo.

Verify After Edits

After the agent edits files, Northstar re-runs typecheck and lint and tells the agent about any new problems it introduced, so it fixes them before finishing. On by default (northstar.verifyOnEdit).

Project Instructions

Put project-specific instructions in .northstar/NORTHSTAR.md, NORTHSTAR.md, AGENTS.md or CLAUDE.md at the workspace root — Northstar reads the first one it finds.

Built-in Guardrails

Northstar ships with a suite of safety checks that run during autonomous execution:

• Loop protector — detects and breaks infinite reasoning cycles
• Git protector — blocks force-pushes and destructive branch operations
• Path guard — restricts file access to your workspace boundary
• Command guard — audits shell commands before execution
• Schema validator — ensures tool outputs are well-formed before injection into context

Tavily API search (18-May-2026)

Northstar now supports Tavily websearch,get a free Tavily api key at https://app.tavily.com

Connect External MCP Servers

Connect Northstar to other MCP servers. Their tools ask for approval unless you're on FULL AUTO or have granted Full Access in the permission settings.

Getting Started

1. Install the extension

Search for Northstar Coding Agent in the VS Code Extensions panel, or install directly from the Visual Studio Marketplace.

2. Open the chat panel

Click the ✦ icon in the Activity Bar, or run Open Northstar Chat from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P).

3. Configure your LLM provider

Click the Settings icon inside the Northstar panel. Set your provider type, base URL, and API key.

Quick start with a free cloud model (default):

Provider : Ollama
URL      : http://localhost:11434/v1
Model    : gpt-oss:120b-cloud
API Key  : (leave blank)

Install Ollama and run ollama signin once — cloud models need a free Ollama account. Selecting Ollama as the provider fills in this model automatically.

Quick start with a fully local model:

Provider : Ollama
URL      : http://localhost:11434/v1
Model    : any model you have pulled, e.g. llama3.2  (ollama pull llama3.2)
API Key  : (leave blank)

Configuration Reference

Basic and common settings are available under → Settings.

Other settings are available under Settings → Extensions → Northstar:

Commands

Requirements

• VS Code 1.109.0 or later
• An LLM endpoint — local (Ollama / llama.cpp) or cloud (OpenAI, Anthropic, OpenRouter)
• No other dependencies required — everything is bundled

Privacy

Northstar sends your messages only to the LLM provider you configure. No telemetry, no analytics, no data is sent to Sugata AI servers. Your codebase stays on your machine if you use local LLMs. You can also use open-source cloud models, or providers such as Anthropic, Google, OpenAI and many others, including custom endpoints.

About Sugata AI

Northstar Coding Agent is built by Sugata AI — an autonomous enterprise AI platform. Our mission is to replace manual workflows with intelligent agentic systems that work alongside your team.