Remote coding control planeChatGPT × Sprites

A real machine behind three familiar tools

Put ChatGPT on areal machine.

zodex gives ChatGPT a Sprite-backed Linux workspace with the exact command, stdin, and patch primitives GPT models already know—then gives you a precise control surface for what can return to GitHub.

Set up zodex See the architecture

03MCP tools

05Rust binaries

∞real shell workflows

ZDX / REMOTE-01 link online

ChatGPTMCP client

zodexdSprite runtime

exec_commandwrite_stdinapply_patch

GitHubPR / grant / YOLO

session / zodexPTY 02:14:31

$cargo test --quiet

✓ checks passed / workspace clean

$zodex-agent github publish-pr --repo amxv/zodex

↗ committed history handed to publisher

POWERFUL SHELLEXPLICIT NETWORK WRITESSTRUCTURAL BOUNDARY

01 / Bring the machine online

The whole setup,
in one screen.

The landing page gives you the main path. The full guide covers the two GitHub Apps, every required permission, verification commands, and troubleshooting.

Open the complete setup guide

01Operator machine

Install zodex

Install the privileged operator CLI on your Mac or Linux machine. This binary never needs to exist inside the agent workspace.

shell

curl -fsSL https://zodex.ashray.xyz/install.sh | sh
zodex --version

02Remote machine

Create a Sprite

Authenticate Sprites, create the Linux workspace, and expose its public URL so ChatGPT can reach the MCP front door.

shell

curl -fsSL https://sprites.dev/install.sh | sh
sprite org auth
sprite create zodex-dev
sprite use zodex-dev
sprite url update --auth public

03Runtime install

Wire the workspace

After creating the reader and writer GitHub Apps, one command installs the runtime, configures Git, starts the services, and smoke-tests the boundary.

shell

zodex sprite setup \
  --sprite zodex-dev \
  --repo owner/repo \
  --reader-app-id <reader-app-id> \
  --reader-pem /path/to/reader.pem \
  --publisher-app-id <writer-app-id> \
  --publisher-pem /path/to/writer.pem \
  --default-base main \
  --url-auth sprite

04ChatGPT

Connect the MCP URL

Add the public HTTPS endpoint as a ChatGPT connector with no additional authentication. The API key already lives in the URL.

shell

https://<sprite-host>/mcp?key=<zodex-api-key>

GitHub Apps are the one deliberate non-command step: a read-only app for clone/fetch and a writer app whose credentials stay behind the publisher boundary. The Quickstart lists the exact settings.

02 / GPT-native by design

Three tools.
Nothing clever in the way.

Frontier GPT models already learned how to code inside a command / stdin / patch harness. zodex preserves that shape and upgrades the environment underneath it.

01Run

exec_command

Start a shell command in a PTY-backed session. Builds, tests, dev servers, REPLs, and ordinary Git all behave like they should on a real Linux machine.

$ cargo test --quiet

02Continue

write_stdin

Poll a live process, send input, or stop it cleanly. Long-running work keeps a stable session handle instead of collapsing into fire-and-forget subprocesses.

session_42 → still running

03Edit

apply_patch

Apply the original Codex-style patch format GPT models already know. The model does not have to relearn a clever abstraction before it can change a file.

*** Update File: src/lib.rs

THE BET

A small API does not mean a small capability surface. It means the model spends its intelligencedoing the work, not translating your agent framework.

03 / Autonomy is a ladder

Start safe.
Move fast when earned.

ChatGPT can clone, edit, test, and commit before any GitHub write window exists. When the work is ready, choose exactly how it leaves the machine.

Compare every write mode

Review first

Publish a PR

ChatGPT commits locally, then asks the isolated publisher daemon to push a generated branch and open a pull request. The shell never sees the writer token.

zodex-agent github publish-pr --repo owner/repo --title "Ship it"

Approve once

Open a push window

Grant one repository a temporary direct-push window. The agent still runs normal Git; zodex enforces the policy underneath the familiar workflow.

zodex-agent github request-push --repo owner/repo

Trusted session

Go scoped YOLO

For repos and tasks you trust, grant bounded autonomy to one repo, several repos, or every installed repo. Grants carry scopes and independent expiry times.

zodex github mode yolo --sprite dev --repo owner/repo --ttl 4h

04 / The trust boundary

The agent cannot flip the switch because the switch is not there.

zodex does not rely on one binary politely checking whether a command is allowed. The privileged operator CLI lives on your machine. The Sprite gets a restricted agent helper. Different binaries, different command surfaces, different credentials.

Read the access model

TRUSTEDyour machine

zodexsetup / upgrades / grants / YOLO

holds operator authority
opens and closes write windows
never installed in the agent shell

RESTRICTEDSprite workspace

zodex-agentstatus / request / publish / revoke

can request, never self-grant
cannot inspect operator credentials
publishes through an isolated daemon

05 / Built for actual work

Everything underneath
the tiny tool surface.

The model sees three simple actions. The runtime handles the operational edges that make those actions trustworthy over hours of real coding.

PTY

A terminal, not a subprocess wrapper

Interactive sessions survive across tool calls, nested processes are cleaned up as a group, and recent error output wins when buffers fill.

GIT

Normal history, normal handoff

Clone, branch, diff, test, and commit exactly as you would locally. Return work through a PR, a temporary grant, or a trusted push window.

BIN

Privilege split by construction

The operator CLI and agent helper are different binaries. The command that can escalate access simply is not installed on the machine ChatGPT controls.

API

One runtime, two front doors

Use the same command, stdin, and patch service through ChatGPT MCP or the direct HTTP API without maintaining two implementations.

06 / Go deeper

The docs follow
the real operating path.

Start with the machine, understand the access model, then operate it like infrastructure—not magic.

Browse every guide

Give ChatGPT a machine
worthy of the model.

Real Linux. Familiar tools. Normal Git. A permissions boundary you can reason about.

Build the workspace View source

Put ChatGPT on areal machine.

The whole setup,
in one screen.

Install zodex

Create a Sprite

Wire the workspace

Connect the MCP URL

Three tools.
Nothing clever in the way.

exec_command

write_stdin

apply_patch

Start safe.
Move fast when earned.

Publish a PR

Open a push window

Go scoped YOLO

The agent cannot flip the switch because the switch is not there.

Everything underneath
the tiny tool surface.

A terminal, not a subprocess wrapper

Normal history, normal handoff

Privilege split by construction

One runtime, two front doors

The docs follow
the real operating path.

Quickstart

Write modes

Architecture

The three tools

Give ChatGPT a machine
worthy of the model.

Put ChatGPT on areal machine.

Install zodex

Create a Sprite

Wire the workspace

Connect the MCP URL

exec_command

write_stdin

apply_patch

Publish a PR

Open a push window

Go scoped YOLO

The agent cannot flip the switch because the switch is not there.

A terminal, not a subprocess wrapper

Normal history, normal handoff

Privilege split by construction

One runtime, two front doors

Quickstart

Write modes

Architecture

The three tools

Give ChatGPT a machineworthy of the model.

Give ChatGPT a machine
worthy of the model.