LocalNest v0.9.17 Get the extension

v0.9.17 · Ollama-only · No API key · No account

AI that reads your project first. And changes nothing until you say so.

LocalNest is a local-first AI coding agent in the Antigravity sidebar. It runs on models you pull with Ollama on your own computer — so your code never leaves your machine, and there is nothing to sign up for.

Get LocalNest View source on GitHub

To get going: install Ollama → ollama pull llama3.2:3b → open the LocalNest panel.

Also known as Vyntra. Inside the editor the panel and commands use the name “Vyntra Chat” — it is the same extension, by the same author.

Chat LOCAL

● Ollama onlinellama3.2:3b · 4 workspace files in context

Changes 0 Terminal 0 Ready

A preview of the panel. Try the tabs — each mode answers differently. Real answers come from your own Ollama model, inside Antigravity.

No cloud modelOnly the Ollama host you choose, localhost by default.
Workspace firstIt indexes your files and picks the matching ones before answering.
Approval before writesProposals are previewed. Nothing is written until you approve it.
Undo built inOriginal files are snapshotted, so an approved batch can be restored.
01 — HOW IT WORKS

Context before code, every single time

LocalNest follows the same order for every question and every task. The model is never handed your question on its own — it is handed your question plus the files that match it.

  1. 1

    Open a project and ask

    Pick a mode — Ask, Plan, Agent, or Build — and type your question or task in the sidebar panel.

    You can also select code and ask LocalNest to explain it, propose a diff-reviewed edit, or fix it from its diagnostics.

  2. 2

    It inspects your workspace first

    Before it answers, it searches your indexed project files and picks the ones that match your question.

    That context travels with the question to the model, so the answer is about your code — not generic code. In Fast mode it uses fewer files, so the first words arrive sooner.

  3. 3

    Ask answers, Plan plans, Agent proposes

    Ask never touches a file. Plan writes only a plan. Agent prepares complete files as one proposal card — and writes nothing yet.

    A plan lists the goal, the exact files to change, the risks, and how to check it worked. Happy with it? “Build this plan” moves it into Agent mode.

  4. 4

    You review every change

    Each proposed file has a preview. Open it beside the editor, read the full version, then approve files one by one, approve them all, or discard the proposal.

    Commands — installing packages, running tests — run only after you press “Run approved commands”. Dangerous commands are refused even if proposed.

  5. 5

    It verifies after you approve

    Once files are approved, LocalNest checks them, runs the checks, and shows the output in the panel.

    If something is still missing or broken, it proposes a fix through the same review and approval — repeating the verify-and-fix loop for up to 5 rounds.

  6. 6

    You can always undo

    The original contents of every file are recorded before approval. “Undo last approved changes” restores edited files and deletes newly created ones from that batch.

    Build mode works the same way for whole apps: describe a website or multi-file project, review the complete proposal, and only then let it be written.

02 — MODES

Four modes, one rule: no surprises

The mode decides what LocalNest is allowed to do with your answer — and, just as importantly, what it is not allowed to do.

AskQuestions, explanations

It does

Answers in the panel, grounded in the workspace files it found. Still searches your project even if no model is reachable.

It never

Changes a file. Ask is read-only by design.

PlanThink before building

It does

Writes only a plan: goal, exact files to change, risks, and how to verify the result.

It never

Changes a file. A plan becomes work only when you choose “Build this plan”.

AgentMulti-file changes

It does

Prepares a complete proposal — new and updated files in one card — then verifies the result after you approve.

It never

Writes anything before approval, or runs a command you have not approved.

BuildWhole apps

It does

Generates the complete files for a website, app, or larger multi-file project as one reviewable proposal. With no model reachable, it still creates a working app from a built-in template.

It never

Writes files without your approval.

03 — REVIEW FIRST

Every change arrives as a proposal, not a fait accompli

Try it: approve the proposal below, then undo it. That is exactly the control you keep inside Antigravity — including the snapshot that makes undo possible.

A proposal is a complete, readable set of files — not a patch you have to reconstruct in your head. You see what will be created, what will change, and what commands it wants to run, before any of it happens.

  • Preview before approvalPress View to open any proposed file beside the editor and read it in full.

  • Approve file by file, or all at onceKeep the parts you want. Discard the rest. Nothing is partial-written behind your back.

  • Commands need a second yesApproved commands run only after you press “Run approved commands” — and dangerous ones are refused outright.

  • Snapshot, then undoOriginal contents are recorded before approval, so one Undo restores the whole batch.

Proposal2 files · Agent mode Awaiting review
src/server.pyUpdated
 def handle_request(req):
+    if not req.user:
+        return error(401, "Sign in first")
     return process(req)
tests/test_auth.pyNew file
+def test_anonymous_request_is_rejected():
+    assert handle_request(anon()).status == 401

Nothing has been written yet. This preview is the whole change.

04 — SETUP

Three steps, then chat

LocalNest does not install a model for you, and it does not download Ollama itself. It checks what is already on your machine and tells you plainly what is missing.

  1. 1Install Ollama

    Download Ollama for your system and run the installer. On Windows, restart Antigravity afterwards so the new terminal picks Ollama up. Choose “use Ollama locally” — no account is needed.

  2. 2Pull a model

    Start Ollama, then pull whichever model you want. A small model is the fastest way to get a first answer; a coding model is the better choice for larger changes.

  3. 3Open the panel and pick it

    Open LocalNest from the nest icon in the activity bar, or the Command Palette. The panel lists the models actually pulled on your machine — pick one and chat. Picking an installed model opens chat straight away.

Terminal — copy any command
$ ollama serveStarts Ollama on http://localhost:11434
$ ollama pull llama3.2:3bSmall and fast — a good first model (about 2 GB)
$ ollama pull qwen2.5-coder:7bCoding-focused — slower on CPU, stronger on code
$ ollama listShows the models already downloaded on this machine

What LocalNest checks on first open

It reads the system: is the ollama program there, is its server running, and which models are pulled? The panel re-checks every time you open it.

  • Ollama is not installed“Install Ollama to continue”, with a button to the download page. No chat is faked in the meantime.
  • Installed, but not runningLocalNest tries to start an already-installed Ollama itself (auto-start is on by default). If that fails, it says to run ollama serve.
  • Running, but no model pulledIt stays on the model picker: pick a model, press Download, and chat opens when the download finishes.
  • A model is already thereIt lists the models found and opens chat automatically — no setup step to get stuck on.

Honest expectations on a laptop

Ollama runs on CPU by default — no GPU setup is needed. The first answer after a download is the slowest, because the model has to load into memory. LocalNest keeps it loaded for 30 minutes, caps the context window, and uses all CPU threads, so later answers come sooner. On a laptop without a GPU, a 3B model will still take tens of seconds for a full answer. That is local CPU, not a failure — and Fast mode, with its smaller context, gets the first words out sooner.

05 — MODELS

Your model, your choice

There is no fixed model and no model list to subscribe to. If Ollama can pull it, LocalNest can use it.

llama3.2:3b

Start here for speed. Small, general-purpose, and the fastest route to a working chat on a laptop without a GPU.

About 2 GB to download. First answer is slowest; later answers speed up while the model stays loaded.

qwen2.5-coder:7b

Choose this for coding. A coding-focused model, better suited to larger edits, plans, and agent tasks.

Heavier on CPU than a 3B model — expect slower answers in exchange for stronger code.

Any Ollama model

Nothing is locked in. Type the exact model name in the panel, or pick it from the list of models pulled on your machine.

Switching models is just picking another one you have already pulled.

You can also point LocalNest at your own remote Ollama host instead of localhost, in Settings. Code then goes only to that host — still no cloud provider, no API key.

06 — QUESTIONS

Asked, answered plainly

No. It sends code only to the Ollama host you configure — http://localhost:11434 by default. There is no cloud model, no API key, and no account. If you point it at a remote Ollama host of your own, code goes only there.

No. Ollama runs on CPU by default and LocalNest has no GPU or CUDA settings. A GPU would make answers faster, but everything works without one — just more slowly, especially on the first answer.

LocalNest says so, plainly, instead of inventing an answer. Ask still searches your workspace and shows what it found; Build still creates a working app from a built-in template. Once Ollama is running and a model is pulled, the same question gets a model answer.

LocalNest is the project name; Vyntra is the display name used inside the editor — for the panel, the commands, and the extension listing. The extension and the author, NitinKumar Hiremath, are the same either way.

No. Files are written only after you approve a proposal, and commands run only after a separate approval. Dangerous commands are refused even when proposed, and every approved batch is snapshotted so it can be undone.

Yes. LocalNest is free, Ollama is free, and the models are free to pull and run locally. The only costs are the download size and the time your own machine spends answering.

Get LocalNest in your sidebar

Install the extension, pull a model, and ask your first question about your own project. If Ollama is not on your machine yet, install that first — LocalNest will tell you, and wait, rather than pretend.

Install Ollama first Source on GitHub

Install it from a VSIX file via Extensions → Install from VSIX. The source, and the latest builds, are on GitHub.

  1. 1. Install Ollama — from ollama.com/download, then restart Antigravity.
  2. 2. Pull a model — ollama pull llama3.2:3b for speed, or a coding model for code.
  3. 3. Open LocalNest — nest icon in the activity bar, or Command Palette → LocalNest: Open Chat.
  4. 4. Pick your model — chat opens when the panel finds a model already pulled on your machine.