Windows 11 · seven days free
A coding agentthat weighsalmost nothing.
A native desktop app and a CLI for working on code with any AI model, built in Rust to stay out of your way: 8.3 MiB idle,1 31 ms to the prompt,2 604 tokens of instructions.3
- Build
- 0.1
- Runs on
- Windows 11, x64
- Made of
- Desktop app + CLI, one Rust engine
- Models
- Your keys, your plan, or local
- Price
- $20 a month after a 7-day trial
Sylphyx
What should we build?
Connected to anthropic/claude-sonnet-5-5. Describe a task, or start from one of these.
acme-web mainChanges the model makes wait here, one hunk at a time, until you keep them or roll them back.
Click anything. Type in the composer, try the terminal (git status, npm test, claude), switch models and watch effort change with them.
01 Measured
Light, by measurement.
Every figure here comes from a script in the Sylphyx repository, run on one reference machine, with the raw samples kept. A build has to pass these gates to ship. Where we have not measured something yet, we say so.
On the scale with the others.
Memory each CLI agent holds at its prompt, doing nothing. A ball's area is its memory: the heavier the agent, the bigger it lands.4
380×the memory OpenCode holds, next to Sylphyx's 0.8 MiB
- Sylphyx0.8 MiB
- jcode~10.4 MB · 12×
- Codex CLI~21.6 MB · 26×
- pi~76.5 MB · 91×
- Antigravity CLI~86.4 MB · 103×
- Cursor Agent~157.5 MB · 188×
- GitHub Copilot CLI~158.1 MB · 188×
- Claude Code~212.7 MB · 254×
- OpenCode~318.4 MB · 380×
- Private working set summed over the app's processes, p95 of each launch, worst of three launches, 1400×860 window, the build of 13 October 2026. Later builds are not measured yet; the limit is a twentieth of the lighter of two desktop AI apps we baselined.
- The CLI build of 15 October 2026. Start-up is the p95 of 50 warm launches, from process start to a prompt that takes keys with the session store open; the first run took 40.2 ms. The memory limit is the lightest CLI agent we measured at its first screen.
- The system prompt and the four tool schemas as the CLI prints them, counted with OpenAI's o200k_base tokenizer: 584 to 604 depending on the wire format.
- Sylphyx is the private working set from note 2; counted with the memory it shares with Windows, it is 6 to 7 MiB. The other CLIs' figures are approximate idle memory at their prompt, not taken by our script, so the methods may differ and the ratios are rough.
Reference machine Intel Core i5-12600KF · 31.8 GiB · RTX 4060 Ti · Windows 11 Pro 26200 · 1920×1080 at 143 Hz
02 Context
604 tokens of instructions. The rest is yours.
Before your first word, an agent sends its own instructions and the definitions of its tools. Sylphyx keeps that under 800 tokens: a short system prompt and four tools. Edits are anchored to the lines the model read, so it rewrites those lines instead of whole files. Your project's AGENTS.md comes after, and its size is yours to choose.
Per-tool counts in OpenAI's Responses format; the chat format wraps them in a few more, which is the 10 at the end. System prompt: 239 tokens.
SylphyxOne request, before your first word
- Total
- 604
- Limit
- 800
- Left for you
- 196
AGENTS.md and your own words are not on this bill.
03 Workbench
Everything beside the conversation.
One window: the session in the middle, and whatever it needs next to it.
- ASessions with their own settings
Each session keeps its own name, model, mode, effort and instructions. Pin it, fork it, export it as Markdown.
- BReview before anything stays
Every edit waits as a hunk you accept or roll back. A checkpoint comes first, so a rollback restores the file exactly.
- CA real terminal
Your own shell, through the same pseudo-console Windows Terminal uses. Run Claude Code, Codex or Gemini CLI right there.
- DA browser beside the chat
WebView2 in a panel: your dev server on localhost, or the docs, without leaving the session. Two panels side by side.
04 Models
Any model. Your keys, or your plan.
Sessions talk to the provider straight from your computer. Subscriptions are used only where their makers allow it.
ClaudekeyGPTkeyGeminikeyGrokkeyDeepSeekkeyMistralkeyGroqkeyOpenRouterkeyChatGPT Plus & ProsoonClaude Pro & MaxplanOllamalocalLM StudiolocalCodexappGemini CLIappCopilot CLIappClaude Codeapp
Bring a key from any of eight providers, a plan where its maker allows it, a model on your own machine, or the makers' own apps in the terminal.
Effort follows the model.
Turn the wheel. The control shows only the levels the model takes, and steps aside for one that takes none.
gpt-5 takes 3 levels; this one is sent as high.
05 Terminal
The same engine, in any terminal.
sylphyx opens a boxed prompt with history, completions and @ mentions. Shift Tab changes the mode, Ctrl C stops a turn, and -p takes one task and finishes it on its own. It idles in under a megabyte.
Click into the terminal and type: /help, /effort, /context, /sessions, /diff.
◆ sylphyx 0.1.0
acme-web
C:\code\acme-web
ask · anthropic/claude-sonnet-5-5
/help for commands · @ mentions a file · shift+tab changes the mode
▸ ask · anthropic/claude-sonnet-5-5 · ▰▰▰▱▱ high · ctx 1.2k · $0.03 · shift+tab mode · /help
06 Privacy
Your code stays on your machine.
Sylphyx runs on your computer and sends your work only to the model you chose. We never see your code, your sessions or your keys.
- Keys you give it live in the Windows Credential Manager, never in a file, a setting or a log.
- Sessions are a SQLite file in your own profile. Export one as Markdown whenever you like.
- Your account holds your name, email, plan and the devices you are signed in on. Nothing about your work.
- Permissions are yours: plan, ask, auto-edit or full auto, set per session.
07 Price
One plan. Both apps.
No tiers to compare and no seats to count. You pay your model provider yourself, or bring a plan that allows it, and Sylphyx never marks up a token.
$20a month
after a 7-day trial of everything
- The desktop app and the CLI, on one subscription
- Every provider, your own keys, local models
- Terminal, browser, review and GitHub inside the app
- Headless runs for scripts and CI
- Every update while you subscribe
Billing isn't live yet, so nothing is charged. When a subscription lapses, the apps still open, export and roll back what you have; they stop starting new turns. Prices in US dollars.
08 Progress
Built in the open, one gate at a time.
The history of main, newest first, with the repository's own commit hashes. Dates are 2026.
- Billing, and a signed installer
What it takes to put the build in your hands.
- Measure what isn't measured yet
- Memory during a live turn, and frame times in a real window
- Tokens per solved task, against another agent on the same frontier model
- Requests to Anthropic's and OpenAI's own endpoints
- Windows 10, macOS and Linux
- Stars, the old effort slider, the mark in the CLI
A warp field behind every new session, the effort control the way Claude's apps drew it, and the Sylphyx mark on the CLI's first screen.
- Accounts and this page
Sign-up with email confirmation and password reset, served by the same small Rust service as the API.
- A workbench around the chat
Sessions with their own settings, effort per model, a real terminal, a browser panel and GitHub.
- The interactive CLI, and a first task solved
A boxed prompt with history and completions. A local 9B model fixed a failing test end to end, three runs out of three.
- An engine with four tools
Sessions, checkpoints, permissions and both wire formats. The CLI passes its start-up, memory and token gates.
- The shell is chosen by measurement
Five prototypes measured. Slint on the CPU idled in 7.3 MiB; the web-view version took 78.6 MiB in seven processes.
Nothing on this page claims more than the measurements show.
09 Questions
Asked so far.
Windows 11 comes first. Create an account and we'll email you the moment your build is ready. macOS and Linux follow.
ChatGPT Plus and Pro work through Sign in with ChatGPT once OpenAI approves Sylphyx; the code is in place. Anthropic's terms don't let other apps use Claude plans, so Sylphyx opens Claude Code in its terminal, where your plan works as usual, and its own sessions use an API key.
No. Requests go from your computer straight to the provider you chose. The account knows your name, email, plan and signed-in devices.
Everything, for seven days, counted from when you start using the app, not from when you create an account here.
Native Rust drawn on the CPU instead of a bundled browser, an engine shared by both apps, and four tools instead of forty. Each choice was measured before it was made.
Lighter than air.
Try it free for seven days. Bring a key, a plan or a local model.