LocalPilot
Free and open source · works offline · your code goes only to the model servers you choose

Cut your AI coding bill to zero.
The agent runs on your laptop.

LocalPilot is a coding agent for VS Code that uses the models you already run with Ollama, LM Studio, llama.cpp, vLLM or Docker Model Runner. It reads your project, makes the edit, runs the tests and shows you the diff. There is no subscription, no account to create and no telemetry to opt out of.

Install from VS Code Marketplace Get started Star on GitHub
v1.1.1 · code --install-extension localpilot.localpilot-agent · or download the VSIX · VS Code 1.104+ · macOS, Linux, Windows
LocalPilot inside VS Code: the agent read src/cart.js, fixed applyDiscount with a +1/−1 edit, ran npm test in the LocalPilot Agent terminal (2 passing) and offers Accept / Reject for the changed file
Captured from a normal session: qwen2.5-coder:7b in Ollama fixed a failing test, ran the suite and asked for review.
Works with Ollama LM Studio llama.cpp vLLM SGLang Docker Model Runner LocalAI · Jan any OpenAI-compatible API
See it work

What it looks like in practice

These are screenshots of LocalPilot v1.1.1 in VS Code with qwen2.5-coder:7b on Ollama. Nothing is mocked up, and you can reproduce each one on your own machine.

Side-by-side diff of the agent's edit to cart.js opened from the Changes panel, with Accept and Reject per file

You see every change before it lands

LocalPilot keeps a copy of each file before it edits it. Open the Changes panel for a side-by-side diff, then accept or reject file by file. If a whole run went wrong, roll the chat back to an earlier checkpoint.

Three agents running at once: tab strip with spinners, Agents panel listing each with its current step, status bar showing 3 agents

Three tasks, three agents, one window

Separate your tasks with | and LocalPilot starts an agent for each one in its own tab. The Agents panel shows what each is doing. When they finish, keep or undo their changes independently.

The LocalPilot Agent terminal showing npm test failing, then passing after the agent's fix, with exit codes captured in the chat

Commands run where you can watch them

When the agent runs npm test or git status, it does so in a LocalPilot Agent terminal inside VS Code. You see the output live and the agent gets the real exit code. Right-click any terminal to have its output explained or a failed command fixed.

The localpilot CLI in a terminal: edits math.js, runs npm test, prints a summary and the changed file

The same agent from the command line

localpilot "fix the failing tests" gives you the identical agent in your shell, with the same tools, skills, memory and approvals. Answer y, n or always at each prompt, or pass --yes in scripts and CI.

Settings → Models: hardware profile (36 GB RAM, Apple M3 Pro, 25 GB model budget), runtimes Ollama / Docker Model Runner / LM Studio, installed models with fits badges and a default marker

It knows what your machine can run

LocalPilot reads your RAM, CPU and GPU memory and turns that into a model budget. Installed models are marked fits, tight or too big. New ones install from a short catalog straight into Ollama, Docker Model Runner or LM Studio.

First-run setup: LocalPilot found Ollama (2 models), Docker Model Runner (2 models) and a Docker-hosted Ollama container; LM Studio, llama.cpp and vLLM not running

Setup starts with what you already have

On first launch LocalPilot checks the usual ports and Docker containers, lists the models it found and suggests one for agent work. If there is nothing to find, one click installs Ollama and a starter model that fits your hardware.

Close-ups

The @ context menu: Files, Active file / selection, @workspace, @selection, @problems, @git, @editors
@ context. Attach files, the current selection, problems, the git diff or your open editors.
Mode picker: Agent (edit files, run commands, verify), Ask (read-only), Plan (investigate and propose a plan)
Agent, Ask and Plan. Ask and Plan can only read; they have no tools that change anything.
Model picker grouped by provider: Ollama, Docker, Docker Ollama container, each model with context size and tools badge
Model picker. Every server LocalPilot found, with context size and tool support, selectable per chat.
Features

What's in the box

LocalPilot was written for small, local models. It falls back to a text protocol when a model cannot call tools, trims old context so 8k windows still finish long tasks, and asks before doing anything you have marked as risky.

Agent

  • Agent modeReads and searches your project, edits files, runs commands, checks diagnostics and retries when something fails. Asks first for risky actions.
  • Parallel agentsRun several tasks at once. Each agent has its own tab, status and set of changes.
  • Review, keep or undoEvery edit is saved first. Diff, keep or revert per file or per session; changed lines are highlighted in the editor.
  • Skills & subagentsSKILL.md playbooks load when they are needed. Nine built-in subagents (explore, coder, reviewer, researcher, clarifier, planner, tester, debugger, docs), or write your own.

Models & setup

  • One-click model setupIf no model is found, LocalPilot installs or starts Ollama and pulls a small starter set that fits your RAM and GPU.
  • Model manager & catalogSee what fits your machine, enable or disable models, set a default, remove them, or install from a curated list with a progress bar.
  • Automatic routingImages go to a vision model, completions to the smallest FIM model, Plan mode to a reasoning model. If the chosen model is offline, the next best one is used.
  • Zero-config, Docker-readyFinds Ollama, LM Studio, llama.cpp, vLLM, Docker Model Runner and model servers inside Docker containers without any settings.

Context & tools

  • Large-context managementEach model gets a token budget. Old tool output goes first, then older turns are summarised, so an 8k-context model still finishes long tasks.
  • Screenshots, images & browser viewCapture a selection, window, screen or web page and let a vision model fix what it sees. A built-in browser view checks your dev server through DevTools.
  • Terminal & CLICommands run in a visible terminal with real exit codes. The localpilot CLI brings the full agent to any shell.
  • Web search & MCPDuckDuckGo, SearXNG, Brave or Tavily search, plus page fetching. Add an MCP server and its tools show up on their own.

Editor & privacy

  • Inline completionsGhost text from fill-in-the-middle models such as qwen2.5-coder, codestral or starcoder2. Works offline.
  • Deep VS Code integrationChat in the sidebar or an editor tab, @localpilot in the built-in Chat view, local models in Copilot's model picker, light-bulb fixes and commit messages.
  • Follows your rulesReads AGENTS.md, CLAUDE.md, .cursorrules and copilot-instructions.md as they are. Nothing to convert.
  • Private by defaultNo account, no telemetry. The extension only talks to the model servers you list, and the defaults are all localhost.
How it works

Up and running in three steps

Install LocalPilot

From the Marketplace, or download the VSIX from GitHub releases.

code --install-extension localpilot.localpilot-agent

Get a model

Click Set up automatically and LocalPilot installs Ollama and a starter model that fits your machine. Already running a server? It will be found on its own.

ollama pull qwen2.5-coder:7b   # or let LocalPilot do it

Ask for a change

Open the sidebar, describe the change, read the diff, keep it. Larger models can be added later from Settings → Models.

⌘⇧L  open chat
⌘⌥K  inline edit
Cost

What you stop paying for

Cloud assistants bill per developer, per month, and meter requests on top. LocalPilot runs on hardware you already own. After the model is downloaded, every prompt, edit and test run is free and works without an internet connection.

Per developerLocalPilotGitHub Copilot Pro / Pro+Cursor ProClaude Pro / Max
Per month$0$10 / $39$20$20 / $100–200
Per year$0$120 / $468$240$240 / $1,200–2,400
Team of 10, per year$0$1,200 / $4,680$2,400$2,400 / $12,000–24,000
Usage limits, premium-request meteringNone. Your hardware is the limitYesYesYes
Needs internet to workNoYesYesYes
Where your code is processedYour laptopVendor cloudVendor cloudVendor cloud
  • Nothing to cancel. Models are files on your disk.
  • No seat licences, trials or request quotas.
  • No vendor account, so there is nothing to leak.

Published list prices for individual plans in USD as of October 2026. Vendors change plans often, so check their pricing pages. Your cost with LocalPilot is electricity. A 7B coding model needs about 8 GB of RAM; 14B to 32B models want 16 to 32 GB.

Compare

How it compares

The same kind of agent workflow as the paid tools, minus the subscription, the rate limits and the upload.

LocalPilotGitHub CopilotCursorClaude Code
Runs fully offline / on-premYesNoNoNo
Bring any local model (Ollama, LM Studio, vLLM…)YesLimitedLimitedNo
Agent mode with tools & approvalsYesYesYesYes
Parallel agents with per-agent change reviewYesPartialYesYes
Skills & custom subagentsYesYesRules onlyYes
Screenshots / vision inputYesYesYesYes
MCP serversYesYesYesYes
Works inside VS CodeYesYesOwn IDEYes
PriceFree · MITSubscriptionSubscriptionUsage-based
Telemetry / account requiredNoneYesYesYes

Based on publicly documented features as of October 2026. These products change often.

Guides

Guides for your setup

Ollama in VS Code

Turn Ollama into a Copilot-style agent inside VS Code with LocalPilot — tool calling, file edits, terminal commands, parallel agents and ...

LM Studio in VS Code

Connect LM Studio's local server to VS Code with LocalPilot and get an agent that edits files, runs commands and reviews changes — using ...

GitHub Copilot alternative

LocalPilot is a free GitHub Copilot and Cursor alternative for VS Code that runs entirely on your hardware with Ollama, LM Studio, llama....

Offline AI coding assistant

How to run a fully offline AI coding assistant in VS Code with LocalPilot — no internet, no accounts, no telemetry. Works on air-gapped m...
Trust

Easy to check

Don't take a landing page's word for it. Three checks, under a minute:

  1. 1
    Turn off Wi‑Fi. LocalPilot keeps answering, editing and running tests.
  2. 2
    Open Settings → Providers. Every address is localhost unless you added one.
  3. 3
    Read the source. It is MIT-licensed on GitHub, and npm test runs against your own Ollama.

Open source

All of it is on GitHub under the MIT license, maintained by . The extension and the localpilot CLI are built from the same code. Each release includes a VSIX and SHA-256 checksums.

It only talks to servers you list

There is no analytics SDK, no account and no update check. The only outbound traffic goes to your model servers, your MCP servers and the web tools you approve.

Tested against a real model

The test suite runs in a real VS Code with a live Ollama. It has the agent read a file, edit it, run the tests, work in parallel sessions and stop when asked.

You set the approvals

Terminal, file edits, deletes and network access each have an Ask, Allow or Block policy, with Careful, Balanced and Autonomous presets. Edits can always be undone.

Your existing rule files work

LocalPilot reads AGENTS.md, CLAUDE.md, .cursorrules, .cursor/rules, .github/copilot-instructions.md, .claude/skills and .claude/agents as they are.

Small models have limits

A 7B model is not as capable as the big cloud models. LocalPilot works around that with retries, a text fallback for tool calls, diagnostics loops and a nudge to finish, and it shows every step so you can step in.

FAQ

Common questions

Which models work best?

For agent work with tools, start with qwen2.5-coder at 7B, 14B or 32B, or qwen3, devstral, gpt-oss or mistral-small. For inline completions use qwen2.5-coder:1.5b, codestral or starcoder2. For screenshots use qwen2.5vl or gemma3. The catalog in Settings → Models marks each size as fits, tight or too big for your hardware and stars a recommendation.

I have never run a local model. How hard is the setup?

One click. When LocalPilot finds no model it offers Set up automatically. That installs Ollama (through Homebrew, winget or the official script), downloads a coder model that fits your RAM and GPU, and configures itself. It picks a smaller model first so you can start within minutes; bigger ones are a deliberate upgrade from the catalog.

Does it use one model for everything?

No. It picks per task: a vision model when the turn has images, the smallest fill-in-the-middle model for ghost-text completions, an optional reasoning model in Plan mode, and the best available model when your default is offline or disabled. You can pin any of these in Settings → Models → Routing.

My model doesn't support tool calling. Does agent mode still work?

Yes. When a model has no native tool support, LocalPilot switches to a text protocol and reads the tool calls out of the model's reply. Results depend on the model, but 7B coder models handle it well.

Does any data leave my machine?

Only if you set it up that way, for example by adding a remote OpenAI-compatible endpoint or approving the optional web search tool. There is no telemetry and no account.

How is this different from Copilot's "bring your own key" option?

Bring-your-own-key swaps the model behind Copilot. LocalPilot is a complete agent: the loop, the tools, parallel agents, change review, skills, subagents, screenshots and MCP. It also registers your local models with VS Code's Language Model API, so other extensions can use them too.

Can I use my existing Claude Code or Cursor rules and skills?

Yes. CLAUDE.md, AGENTS.md, .cursorrules, .cursor/rules/*.mdc, .claude/skills, .claude/agents and .github/copilot-instructions.md are all read without any changes.

Is it on the Marketplace?

Yes. Search for LocalPilot Agent in VS Code or install localpilot.localpilot-agent. Each release is also published as a VSIX on GitHub Releases with SHA-256 checksums.

Try it on your next bug fix.

Free, MIT-licensed, no account. Install it, point it at Ollama, and keep your code at home.

Install LocalPilot Read the docs