SCODE

Fast, persistent coding CLI for AI workflows — optimized for local models and execution-heavy coding tasks.

SCODE is a free and open-source terminal runtime for coding work. It can read and edit files, run commands, keep sessions alive, connect MCP tools, and work with cloud or local models.

macOS · Windows · Linux

Install

Requires Node.js 24+.

npm install -g https://sidra-ai.com/scode/scode-1.0.2.tgz

Then run:

scode

That's it.

Quick use

Open the terminal UI:

scode

Run one prompt:

scode -p "inspect this project and explain the architecture"

Run in another workspace:

scode --cwd /path/to/project

Continue the latest session for the current directory:

scode --continue

Useful commands inside SCODE:

/model
/effort
/help
/compact
/new
/resume
/rewind

Why SCODE exists

SCODE started from the ZCode architecture.

We kept the parts that make ZCode useful as a coding runtime — the execution engine, session model, terminal UI, MCP support, file/edit/search tools, and provider architecture — then removed or reduced layers and flows that were unnecessary for SCODE's narrower goal: fast, persistent execution with less coordination and context overhead.

The result is a leaner execution runtime designed to stay close to the workspace and let the reasoning layer remain replaceable.

That matters especially for local models, where unnecessary context, orchestration, and repeated tool-round trips can cost far more than they do with large cloud models.

SCODE is not a wrapper around ZCode. It is an independent product derived from it, with its own runtime behavior, local-model path, persistent execution model, cross-platform packaging, and SIDRA integration.

Upstream attribution and third-party license material are preserved in NOTICE.md and THIRD-PARTY-NOTICES.md.

Built for local models

SCODE is designed to work well with smaller and local coding models, not only large hosted models.

On Apple Silicon, SCODE supports an MLX local-model path. On macOS, Windows, and Linux, it can also discover local Ollama models automatically.

Start SCODE and use:

/model

to select an available model.

SCODE keeps local inference optional:

• use a cloud model if you want the strongest reasoning layer
• use MLX on Apple Silicon for fast local inference
• use Ollama for a cross-platform local-model path
• keep a persistent SCODE runtime alive so repeated work does not need to rebuild the execution environment from scratch

A controlled SCODE benchmark also showed why fresh, bounded execution context matters for local models: the same warm Qwen3.5-4B local workflow completed in about 3.88s on a fresh execution session versus about 35.95s on an accumulated session, while input/context volume dropped by roughly 70%.

That is a SCODE session-design benchmark, not a ZCode-vs-SCODE benchmark. The point is that SCODE is deliberately shaped to minimize execution and context overhead where it matters most.

Persistent execution

SCODE can run as a normal CLI, but it can also act as a persistent execution cabinet behind another AI client.

Instead of repeatedly starting a new coding process, a client can attach to one workspace-scoped SCODE runtime and reuse the same execution session across many operations.

This lets one external request fan out into many internal operations such as:

Read
Edit
Search
Code Index
Shell
MCP
Session state

For multi-step coding tasks, this reduces external tool round-trips and keeps execution state close to the project.

Architecture

AI / Model
    |
    v
  SCODE
Persistent Coding Runtime
    |
    +-- Read / Edit / Search
    +-- Shell / Commands
    +-- Code Index
    +-- Sessions / Resume
    +-- MCP tools
    +-- Local models
    |
    v
Your workspace

SCODE does not require SIDRA OS. It is an independent open-source project.

SCODE + SIDRA OS

SCODE focuses on persistent coding execution.

SIDRA OS is the broader execution layer that can connect AI clients to the computer itself: files, terminal, browser, computer control, project memory, system tools, channels, and additional runtimes.

SCODE can connect to SIDRA OS through MCP, so the two can be used together:

ChatGPT / Claude / Local Model
              |
              v
           SCODE
   Persistent Coding Runtime
       /        |         \
  Read/Edit   Shell     Code Index
                   \
                    \\ MCP
                     v
                  SIDRA OS
        Files / Browser / Computer
        Memory / Tools / Channels
        Workspace / System execution

That gives you two clean modes:

SCODE alone — a focused coding runtime.

SCODE + SIDRA OS — coding execution plus a broader computer execution layer.

The integration is optional. SCODE remains fully usable as a standalone CLI.

SIDRA OS repository:
https://github.com/sidra-ai-development/sidra-os

SIDRA OS website:
https://agent.sidra-ai.com/

Start SIDRA OS:
https://agent.sidra-ai.com/start

Why SCODE

• Persistent workspace sessions
• Built-in read, edit, search, shell, and coding tools
• Low-overhead terminal UI
• Persistent code index
• MCP support
• Cloud-model support
• MLX support on Apple Silicon
• Ollama auto-detection on macOS, Windows, and Linux
• Session resume and continuation
• Cross-platform runtime behavior
• Optional integration with SIDRA OS

Build from source

Requirements:

• Node.js 24.14.0
• pnpm 10.33.2

git clone https://github.com/sidra-ai-development/scode.git
cd scode
pnpm install --frozen-lockfile
pnpm build

Run the development CLI:

pnpm dev -- --help

Source layout

apps/scode-cli/
  packages/
    cli/          command-line entry point
    core/         execution and session runtime
    tui/          terminal UI
    bootstrap/    runtime assembly
    adapters/     model, storage, MCP and platform adapters

packages/
  shared/         shared runtime primitives
  provider*/      provider support
  zcode-cua/      computer-use compatibility layer

scripts/          build and release support
third-party/      license and provenance material

Some internal package names still use the upstream @zcode/* namespace for compatibility with the architecture SCODE was derived from. The public product and CLI are SCODE.

License

SCODE is licensed under Apache-2.0.

SCODE contains code derived from ZCode. Upstream and third-party notices are preserved in NOTICE.md and THIRD-PARTY-NOTICES.md.

---

Built by SIDRA AI Development.