THE CODING AGENT. ON YOUR TERMS.

Safe. Speedy.
Self-evolving.

Extraordinary capability.
Uncompromising control.
An open-source coding agent built around you.

Get started Explore the source

Developer Preview · macOS & Linux

sS-CODE ENGINE
Sandboxed execution
Built in Rust
Your model. Your machine.
Open source.Local first.Yours to inspect.

01 / SAFE

Powerful by nature.
Protected by design.

A coding agent should earn your trust.
Start with a Rust core, enforced boundaries,
and a record of what happened.

SANDBOXED COMMANDS

Room to build.
Boundaries that hold.

Built-in commands run inside an OS sandbox. Writes stay scoped to your workspace. Network access starts off.

RUST AT THE CORE

Trust starts
under the hood.

A Rust execution core. Local policy checks. Credentials kept out of the browser. Open source, so you can look closer.

VISIBILITY & CONTROL

Your context.
In clear view.

Review tool requests, approvals, and outcomes in the session record. Understand the work behind the answer.

On the roadmap: one view of every file the model has seen.

Session activityIllustration
↗Read workspace filesrc/app.tsAllowed
⌘Run project testsWorkspace sandboxScoped
⊘Read sensitive file.env.productionBlocked
Policy checked before execution

Model requests go to your chosen endpoint. Trusted host integrations have separate execution boundaries. Explore the security model ↗

02 / SPEEDY

Less repeated work.
More forward motion.

Precise edits. Focused context. A leaner path from task to done.
And measurements you can inspect.

34.7%

fewer reported tokens

6.6%

lower median elapsed time

3/3

runs passed · both agents

Same model. Same task.

Durable Task Queue · GLM 5.3

Median reported total tokens · lower is better

Historical development build · August 31, 2026 · Three runs per agent on one task. Token accounting is harness-reported, not normalized billing. S-Code had the slower worst run (102.896s vs 55.949s). These results do not establish a general speed ranking or describe the current release.

See every result and the methodology ↗

03 / SELF-EVOLVING

Built to learn.
Designed to become more.

Good work should inform what comes next.
Turn feedback into reusable knowledge.
Build toward better skills. And a better harness.

01

Do.

Work on real tasks.

02

Reflect.

Learn from the outcome.

03

Improve.

Carry knowledge forward.

IN THE PREVIEW

Every session can
leave something useful.

Task feedback, bounded retries, and memory you explicitly choose to save. Reusable context for the next session. Repeatable evaluations to measure progress.

THE DIRECTION

Skills that grow.
A harness that evolves.

Our ambition: update skills and the harness through learning, so the agent improves how it works—not just what it remembers.

Autonomous learning and self-upgrades are not implemented in the current public Preview.

s

Your next idea.
Meet your new agent.

Your terminal or your browser.
Your choice of model. One shared workspace.

Install from source
git clone https://github.com/sai-foundation/s-code.git
cd s-code
scripts/install-from-source.sh
./s-code setup
./s-code doctor
./s-code
Prefer a browser? ./s-code webmacOS & Linux
Read the documentation ↗View on GitHub ↗