v1.118.8 Linux x64 and arm64 Windows x64 Linux Windows

The advanced agentic code editor.

Chat, inline edits and an agent that works across your files — on the model you choose, hosted or local. No telemetry, no account, no relay.

DownloadDownload for LinuxDownload for Windows Read the docs

Free to download and use — you bring your own model keys.macOS isn't available yet; Linux and Windows are.All platforms

File
Edit
Selection
View
Go
Run
Terminal
Help

aurora

Explorer

Aurora

src
routes
TShealth.tsM
TSmetrics.ts
lib
TSratelimit.tsM
TSlog.ts
TSserver.ts
test
{}package.json
M↓README.md

Outline

Timeline

TShealth.ts

TSratelimit.ts

TSserver.ts

src routes health.ts ƒ router.get() callback

KeepUndo

1import { Router } from './router';

2import { log } from '../lib/log';

3import { TokenBucket } from '../lib/ratelimit';

5const router = new Router();

6const bucket = new TokenBucket({ capacity: 20, refillPerSecond: 5 });

8router.get('/health', (req, res) => {

9 if (!bucket.take(req.ip)) {

10 log.warn('health: rate limited', { ip: req.ip });

11 return res.status(429).json({ ok: false, retryAfter: 1 });

12 }

13 res.json({ ok: true, uptime: process.uptime() });

14});

16export default router;

Chat

Rate‑limit the health route and log rejected requests.

Sirius

Running search files…
Running read file…
Running read file…

Added a token bucket in front of /health — 20 requests, refilling 5 a second — and a log.warn for every rejected request.

TS health.ts +7
TS ratelimit.ts +14

2 files changedKeep All EditsUndo All Edits

Describe what to build

Agent qwen3:32b

main 0 0

Ln 13, Col 3Spaces: 2UTF-8LFTypeScript qwen3:32b

Bring your own model

Twelve providers, hosted or on your own machine, behind the editor’s own chat, inline edits and agent mode. Keys live in your keyring.

Agentic by default

The agent reads, searches, runs and edits across files. Every change comes back as a diff you keep or undo, with checkpoints behind it.

Collects nothing

No telemetry, no crash reports, no analytics, no account. Requests go straight from your machine to the provider you chose.

01 — Agent mode

An agent that works across your files. You keep the last word.

Describe the change. Sirius reads, searches and edits across files and runs what it needs in the terminal.

Chat

Send a Retry-After header on the 429 and add a test.

Sirius

Running read file…
Running edit file…
Running create file…
Running run in terminal…

Added retryAfter() to the bucket, sent it as a Retry-After header and covered it in health.test.ts. The tests pass.

TS health.ts +3 −1
TS ratelimit.ts +4
TS health.test.ts +14

3 files changedKeep All EditsUndo All Edits

TShealth.ts

TSratelimit.ts

TShealth.test.ts

KeepUndo

8router.get('/health', (req, res) => {

9 if (!bucket.take(req.ip)) {

return res.status(429).end();

10 const wait = bucket.retryAfter();

11 res.set('Retry-After', String(wait));

12 return res.status(429).json({ wait });

13 }

14 res.json({ ok: true });

15});

Diffs, not surprises. Keep or Undo per hunk or per file, with checkpoints to step back to.
Tools that fit the model. Sized in tiers; prompted tools for local models without a tool API.
Ask, Edit or Agent. Your choice per message.

02 — Tab and next-edit

Completion that runs on your machine.

Fill-in-the-middle completion against a local model you choose, Ollama or llama.cpp. Next-edit prediction offers your next change as a Tab-able diff. No code goes to a cloud.

Finds your model. On auto, a running Ollama code model.
Next-edit is opt-in. Off by default, using the same local model.
Quiet by design. Debounced, cached, cancellable.

03 — Bring your own model

Twelve providers. No lock-in.

Hosted or on your own machine, in one picker.

✨ Select AI Model

Choose a model from any provider... (🧠=Thinking 👁️=Vision)

Anthropic Claude
Claude Opus 5 🧠👁️claude-opus-5
Best all-round coding and agentic model — thinking on by default
Claude Sonnet 5 🧠👁️claude-sonnet-5
Strong balance of speed, cost and intelligence — a good daily driver
Google Gemini
Gemini 3.1 Pro 🧠👁️gemini-3.1-pro
Gemini 3.5 Flash 🧠👁️gemini-3.5-flash
Ollama (local) — detected
qwen3:32b qwen3:32b
gemma3:27b gemma3:27b

Switch in the chat. The model picker lists every model you can reach; the status bar shows which is answering.
Keys stay in the keyring. Never settings.json, never synchronised.
No account, no relay. Requests go straight to the provider you chose; a local model never leaves your machine.

Hosted Anthropic ClaudeGoogle GeminiOpenAIOpenRouterGroqDeepSeekMistralxAI Grok

Local OllamaLM Studiollama.cppvLLMAny OpenAI-compatible endpoint

04 — Project rules and context

Your project's rules, in every request.

Put the build commands, the conventions and what never to touch in .siriusrules, .sirius/rules.md or AGENTS.md. Sirius reads it before it answers.

.siriusrules

1# aurora

2pnpm build, pnpm test.

3- Never use npm or yarn.

4- Tests sit beside the code.

5- Never log request bodies.

6- Do not touch generated/.

TShealth.ts

8router.get('/health', (req, res) => {

9 if (!bucket.take(req.ip)) {

10 log.warn('rate limited');

11 return res.sendStatus(429);

Add a test for the rate limiter.

Running read file…
Running create file…
Running run in terminal…

Added ratelimit.test.ts beside ratelimit.ts and ran pnpm test.

TShealth.ts

Describe what to build

Agent qwen3:32b

Context, no ceremony. The file in front of you travels with each request.
AGENTS.md is shared. Other agents read it too.
Rules never block. A missing file is simply skipped.

05 — Integrated browser

The agent can see the page too.

A Playwright-backed browser lives in the editor, and the agent gets six tools for it, on by default.

Read first, look when it helps. read_page returns the accessibility tree; screenshot_page reaches models that accept images.
Real Playwright. No separate Chrome to install.
Offered where it fits. Small local models are not handed schemas they cannot hold.

06 — Git assist

Commit messages and conflicts, from the Source Control view.

Commit messages come from your diff — a summary under 72 characters, then the why. Merge conflicts go to the agent, which explains each resolution as a diff to review.

Source Control

Rate-limit /health with a token bucket

Reject requests over 20 per client, refilling 5 a second, and log each rejection.

Commit

Staged Changes 2

TShealth.tssrc/routesM
TSratelimit.tssrc/libA

Chat

Resolve the merge conflicts in: src/lib/ratelimit.ts. Read each file, understand both sides of every conflict, and apply a resolution that keeps the intent of both changes. Explain each resolution briefly as you go.

Running read file…
Running edit file…

Both sides changed the bucket's options. I kept your capacity of 20 and main's refill rate of 5 a second.

1 file changedKeep All EditsUndo All Edits

Into the commit box. Staged diff first, working tree if nothing is staged.
Any model, including local.
No model, no mystery. It says to add a key or start Ollama.

07 — Import

Switch editors. Bring your setup.

One command brings your settings, keybindings, snippets and extensions from Visual Studio Code, Cursor, Windsurf (now Devin Desktop) or VSCodium.

It finds what is installed. Only editors on this machine are offered.
Your choice, category by category. All four are ticked by default.
A report, not a shrug. Counts, and anything not on Open VSX.

08 — Remote server

The server half, shipped with every release.

The headless half a remote extension installs on the machine you connect to. Unpack it, run bin/sirius-server, and /version answers.

ProblemsOutputDebug ConsoleTerminalPorts

deploy@edge-02:~$ mkdir sirius-server && tar -xzf sirius-server-linux-x64.tar.gz -C sirius-server --strip-components 1

deploy@edge-02:~$ sirius-server/bin/helpers/check-requirements.sh

deploy@edge-02:~$ sirius-server/bin/sirius-server --version

1.118.7

603044c6cfb193c74774c10e45bbef1a02e30682

x64

deploy@edge-02:~$ sirius-server/bin/sirius-server --host 127.0.0.1 --port 8317 --accept-server-license-terms --without-connection-token &

[1] 4128

Server bound to 127.0.0.1:8317 (IPv4)

Extension host agent listening on 8317

deploy@edge-02:~$ curl -s http://127.0.0.1:8317/version; echo

603044c6cfb193c74774c10e45bbef1a02e30682

deploy@edge-02:~$

A floor you can plan on. glibc 2.28: Debian 10, RHEL 8 or newer.
Tested before every release. Started on three distributions, x64 and arm64.
Wired, not yet proven. Remote extensions can fetch it; an end-to-end connect is unproven.

Privacy

Collects nothing.

Sirius has no telemetry to turn off: the build has no telemetry destination, so nothing is ever sent. Beyond the request you send to the provider you picked, the only things that leave your machine are an hourly update check that carries no identifier and the Open VSX queries you make when you browse extensions — and with a local model, the provider request is not even that.

This site holds to the same rule: no third-party scripts, no cookies, no analytics. Read the privacy statement.

No telemetry, no crash reports, no analytics, no account. There is no setting to find: there is no code in Sirius that reports anything about you anywhere.
No relay. A request goes from your machine to the provider you chose. Sirius runs no proxy, so there is nothing for it to log.
Keys in the keyring. Encrypted by the operating system — the Keychain on macOS, DPAPI on Windows, libsecret or KWallet on Linux — never settings.json, never synchronised.
Local means local. With Ollama, LM Studio, llama.cpp or vLLM, no prompt, file or fragment of code reaches a third party.

Platforms · v1.118.8

Linux and Windows today. macOS when it can be signed.

Every build comes from the release pipeline and is attested. The Linux builds are installed and run on clean machines before anything is published; the Windows installer is checked for size, checksum and provenance.

Linux

x64 · arm64

Tarball
.deb for Debian and Ubuntu
.rpm for Fedora and RHEL
The [sirius] pacman repository for Arch (x86_64)
The remote server, x64 and arm64

The editor notifies you of updates; packages update through your package manager.

Windows

x64

Installer, per user or per machine
Updates download in the background and install on restart

The installer is not code-signed yet, so SmartScreen warns on first run: choose More info, then Run anyway. Signing is on the roadmap.

macOS

not yet

The packaging exists, but a macOS release needs an Apple Developer certificate for signing and notarisation; without it Gatekeeper refuses the app.

Get notified on GitHub

All downloads and checksums

Compared, honestly

Where Sirius stands, cell by cell.

Compared on October 8, 2026 from each vendor's own pages; every competitor cell has a recorded source, and a cell without one is left blank rather than guessed.

Capability	Sirius	Cursor	Devin Desktop (formerly Windsurf)	Zed	Google Antigravity	Kiro	Trae
Bring your own model	Twelve providers + any OpenAI-compatible endpoint; keys in your keyring.	OpenAI, Anthropic, Google, Azure, Bedrock keys — requests still pass through Cursor’s servers.†	— (no source recorded)	Anthropic, OpenAI, Google, Mistral, DeepSeek, xAI and compatible endpoints.†	No.†	Models from Kiro’s own list; no key setting documented.†	Custom models by API key.†
Local models	Ollama, LM Studio, llama.cpp, vLLM; Tab completion runs on them.	No — every request goes through Cursor’s servers.†	No — “Devin Local” is a local agent on hosted models.†	Ollama, LM Studio, llama.cpp, any local OpenAI-compatible server.†	Not in the editor; the SDK only.†	— (no source recorded)	— (no source recorded)
Telemetry and relay	None — no telemetry, no account, no relay.	Privacy Mode: zero retention. Off: prompts, code and editor actions may be stored and used for training. Requests pass through Cursor.†	— (no source recorded)	Anonymous telemetry; client-side reporting can be switched off.†	— (no source recorded)	Usage data and content collected by default on Free and individual plans; opt out in settings.†	Chats and code may be used for analytics and training unless Privacy mode is on.†
Price	Free. You pay your provider, or run local models for nothing.	Hobby free · Pro $20/mo · Pro+ $60 · Ultra $200 · Teams from $40/user.†	Free · Pro $20/mo · Max $200/mo · Teams; usage allowances, then API pricing.†	Personal $0 with your own keys · Pro $10/mo · Business $30/seat.†	Individual $0 with weekly limits · Pro $20/mo · Ultra $100 and $200/mo.†	Free 50 credits/mo · Pro $20 · Pro+ $40 · Pro Max $100 · Power $200.†	Free · Pro $20 · Pro+ $60 · Ultra $200 per month.†
Source	Proprietary licence; source published on GitHub; built on Code - OSS (MIT).	Proprietary.†	— (no source recorded)	Open source, GPL-3.0.†	— (no source recorded)	Source not published; built on Code - OSS.†	— (no source recorded)
Platforms	Linux x64 and arm64 (tarball, .deb, .rpm, pacman repo); Windows x64, unsigned; macOS not yet.	macOS, Windows, Linux.†	macOS, Windows, Linux.†	macOS, Windows, Linux.†	macOS, Windows, Linux.†	macOS, Windows, Linux.†	macOS, Windows, Linux (.deb, x64).†
Agent mode	Yes — diffs with checkpoints; tools scale to the model.	Yes.†	Yes (Devin Local).†	Yes (Agent Panel).†	Yes — agent-first.†	Yes — spec-driven.†	Yes.†
Tab completion	Yes, on a local FIM model; next-edit prediction optional.	Yes.†	Yes — unlimited on every plan.†	Yes — Edit Prediction, 2,000/mo free.†	Yes — unlimited on the free tier.†	— (no source recorded)	Yes — 5,000/mo free, unlimited on paid.†
Remote development	Server ships with every release (x64, arm64); a remote-extension connect is not yet proven — we say so.	Remote SSH.†	SSH remoting to Linux hosts; Dev Containers.†	SSH remoting; macOS, Linux and Windows targets.†	— (no source recorded)	Open VSX remote SSH extensions.†	SSH remote development; Linux hosts only.†

† Each marked cell is backed by an entry in comparison.sources.md, with the page it was read from. Competitor names belong to their owners. Tell us where a cell is wrong: open an issue.

FAQ

Questions, answered plainly.

Anything missing? Ask on GitHub.

Is Sirius free?

Yes. Sirius is free to download and use on any number of devices you own, for personal or commercial development. There are no plans and no tiers: you bring your own model keys and pay the provider directly, or run local models and pay nothing.

Which models can I use?

Anthropic Claude, Google Gemini, OpenAI, OpenRouter, Groq, DeepSeek, Mistral and xAI through their APIs; Ollama, LM Studio, llama.cpp and vLLM locally; and any other OpenAI-compatible endpoint. Pick a model with Ctrl+Shift+M or from the picker in the chat.

Does my code leave my machine?

Only in the request you send, and only to the provider you chose: the prompt, the context you attached or the editor included, and files a tool read. Sirius runs no relay. With a local model, nothing leaves the machine at all. There is no telemetry, no crash reporting and no account.

Is this Visual Studio Code?

No. Sirius is a derivative work of Code - OSS, the MIT-licensed open-source foundation of VS Code; it is not VS Code and is neither produced nor endorsed by Microsoft. It does not use Microsoft's branding, telemetry or Marketplace: extensions come from Open VSX, because Microsoft's marketplace terms bar other products from using theirs.

Can I bring my extensions and settings?

Yes. Sirius: Import from VS Code, Cursor, or Windsurf brings your settings, keybindings, snippets and extensions over from Visual Studio Code, Cursor, Windsurf (now Devin Desktop) or VSCodium, and reports what it imported.

Where are my API keys stored?

In the editor's secret storage, encrypted by the operating system — the Keychain on macOS, DPAPI on Windows, libsecret or KWallet on Linux. Never in settings.json, never synchronised, never sent anywhere but the provider they belong to.

How does it update?

Every build checks the update server hourly and sends only the platform, channel and build commit. On Windows the editor downloads the installer in the background and runs it on restart; on Linux it notifies you and opens the download page, and the .deb, .rpm and Arch packages update through your package manager — the pacman repository is updated by the release itself. update.mode: none turns the check off.

Why does Windows warn me on first run?

The Windows installer is not code-signed yet, so SmartScreen shows its warning. Choose More info, then Run anyway. Every release asset is attested with GitHub build provenance and carries a sha256 sidecar you can verify. Signing is on the roadmap.

Is there a macOS build?

Not yet. The packaging exists, but a macOS release needs an Apple Developer certificate for signing and notarisation; without it Gatekeeper refuses the app. Follow the GitHub issue to hear when it ships.

Is Sirius open source?

No. Sirius ships under its own proprietary licence, and its source is published on GitHub. Code - OSS and every third-party component keep their own licences, reproduced in ThirdPartyNotices.txt.

v1.118.8 · Linux and Windows

Install Sirius. Bring a model. Start working.

Install it, add a provider key or start Ollama, and it thinks.

DownloadDownload for LinuxDownload for Windows Getting started

# Arch Linux — the [sirius] pacman repository
sudo pacman -Syu sirius-ide-bin