Roadmap
Where Sage is going, what will never change, and what we have deliberately ruled out. Direction rather than dates — because a date we cannot keep is worth nothing to you.
What this page is, and is not
This is a statement of direction, not a schedule. You will not find dates or version numbers here, because a date we cannot keep is worse than no date at all. What you will find is the set of commitments that constrain every decision we make, and the areas we are actively pushing on. If something below turns out to be wrong for users, we will change it rather than ship it to protect a public promise.
Commitments that do not change
These are not features to be delivered. They are constraints on everything we build, and they are the reason to trust the rest of this page:
•
Local-first stays the default. Running a model on your own hardware will always be the path of least resistance, not a degraded fallback for people who opt out of the cloud.
•
Your code is never training data. Not now, not as a paid opt-out, not buried in a settings toggle.
•
Cloud requests stay anonymized. Identifiers stripped, emails and phone numbers scrubbed, before anything leaves your machine.
•
Deletion stays real. When you delete your account, it is gone immediately — no shadow copies, no grace period during which it still exists.
•
Open weights stay the foundation. Sage is built on models you can inspect, download, and run without us.
•
Sage keeps working offline. If our infrastructure disappeared, your local setup would keep functioning.
Where the work is going
Broadening model coverage. Sage supports over 200 open-weight models across the Llama, Gemma, Qwen, Mistral, Phi, DeepSeek-Coder, and Devstral families. The open-weight ecosystem moves quickly, and keeping current with it — new releases, better quantizations, models that fit on hardware people actually own — is continuous work rather than a milestone.
Making local models more usable on ordinary machines. The gap between what a model can do and what your laptop can hold is the single biggest constraint on local-first AI. Anything that narrows it — smarter memory use, better quantization support, clearer guidance about which models suit which hardware — is high-value work.
Deepening the integrations that already exist. Sage works with VS Code, Cursor, JetBrains, Neovim, Sublime Text, Zed, GitHub CLI, Git Hooks, Godot 4, and the Unity Editor. Our bias is toward making each of those genuinely good rather than adding a long tail of shallow ones.
Sharpening the agentic loop. The CLI reads files, writes files, runs commands, and iterates on test results. The difference between a tool that helps and one that wastes your afternoon is how well it recovers when it gets something wrong — better context handling, better use of /test and /undo, fewer confident mistakes.
Extending the SMS bridge. Firing a task at your own computer from iMessage, Google Messages, or email is a genuinely different way to work. Making it more reliable and easier to secure — clearer device registration, tighter control over who can send — is ongoing.
Making cost legible. You are billed on output tokens, and you should always be able to see exactly where you stand before an overage rather than after. Usage visibility in the app and from /status keeps improving.
Things we are deliberately not doing
A roadmap is more informative for what it rules out than for what it promises:
•
No advertising, no behavioural tracking, no data brokerage. There is no version of Sage where your attention is the product.
•
No cloud-only lock-in. We will not build a feature that works in the cloud and quietly cannot work locally, if a local path is technically possible.
•
No dark patterns in billing. Cancelling stays one page and one click, and we will not make downgrading harder than upgrading.
•
No feature we cannot support. Shipping something impressive that we then abandon costs users more than never shipping it.
How to influence this
Email support@sageworksai.com and tell us what you are trying to do and where Sage gets in the way. Concrete accounts of real friction — the model you were running, the machine you were on, what you expected — move priorities far more than feature requests in the abstract. We would rather fix the thing that wasted your afternoon than build the thing that sounds good in a list.
On honesty
This page will never contain a fabricated date, an invented milestone, or a feature listed as shipped that is not. If something is here, it is a direction we are genuinely working in. If a direction changes, this page changes with it.