ZIPCODE

A private coding agent for your terminal.

Source-level Codex fork · GitHub device login · Self-hosted Qwen3.8 inference

ZIPCODE is a source-level fork of the OpenAI Codex CLI connected to a private, self-hosted model service. It keeps the terminal agent, sandbox, approvals, tools, MCP support, and session workflow while using the real Qwen model identities served by the ZIPCODE team.

Install

macOS and Linux:

curl -fsSL https://raw.githubusercontent.com/NeelM0906/zipcode/main/install.sh | sh

Windows PowerShell:

irm https://raw.githubusercontent.com/NeelM0906/zipcode/main/install.ps1 | iex

The installer downloads the release for your platform, verifies its SHA-256 checksum, and places the executables in your user path. Every release archive also receives signed GitHub/Sigstore build provenance. See the installation guide for version pinning, manual installation, and removal.

Sign in and run

zip-code login
zip-code

zip-code login opens GitHub's device-authorization page. Your GitHub account must have an active ZIPCODE invitation. No OpenAI API key or ChatGPT account is used. Unless tracing is disabled, the first launch displays ZIPCODE's full-trace collection notice and requires explicit acceptance before the coding runtime starts.

To run without recording or uploading coding traces:

export ZIPCODE_DISABLE_TRACE_UPLOAD=1
zip-code

Run a non-interactive task with:

zip-code exec "explain this repository"

Switch models explicitly when a task needs the full context route:

zip-code -m Qwen/Qwen3.8-27B-FP8

The default is Qwen/Qwen3.8-Flash-Next with a 524K configured context. The full route is Qwen/Qwen3.8-27B-FP8 with a 1M configured context. These are serving limits for the current deployment, not generic capability guarantees.

TinyFish web search

Set the TinyFish API key in the environment before starting ZIPCODE:

export TINYFISH_API_KEY="..."

$env:TINYFISH_API_KEY = "..."

Then select TinyFish with live web access in ~/.zipcode/config.toml:

web_search = "live"

[tools.web_search]
provider = "tinyfish"

The key is used only for the fixed TinyFish search endpoint and is removed from model-reachable child-process environments.

What is included

• A Rust zip-code front end with isolated ~/.zipcode state.
• A ZIPCODE-branded Codex runtime built from this repository's source.
• GitHub device authorization, invitation checks, 15-minute access JWTs, rotating refresh tokens, and immediate revocation.
• A two-model Responses-compatible gateway for Qwen3.8 Flash-Next and 27B FP8.
• Reproducible SGLang deployments, health checks, observability, benchmarks, and retained long-context evidence.
• Optional, consent-gated full-fidelity rollout traces uploaded to private Supabase Storage for evaluation and model-training dataset preparation.
• Native release builds for Linux x86_64, macOS Apple Silicon and Intel, and Windows x86_64.

Documentation

• Installation
• Authentication and invitations
• Trace collection and storage
• Configuration
• Sandbox and approvals
• Gateway and deployment report
• Measured optimization results
• Security boundary
• Contributing

Build from source

Install a current Rust toolchain, then build both source-level executables:

git clone https://github.com/NeelM0906/zipcode.git
cd zipcode/codex-rs
ZIPCODE_GITHUB_CLIENT_ID=YOUR_PUBLIC_OAUTH_CLIENT_ID \
  cargo build --release --locked \
  -p codex-cli --bin codex \
  -p codex-zipcode-cli --bin zip-code \
  -p codex-code-mode-host --bin codex-code-mode-host

Install target/release/zip-code as zip-code, install target/release/codex next to it as zip-code-core, and keep codex-code-mode-host in the same directory. Release builds embed only the public GitHub OAuth client ID; signing secrets remain on the gateway.

Operate the private service

The service path is:

zip-code
  -> GitHub device identity + ZIPCODE invitation
  -> authenticated HTTPS edge
  -> ZIPCODE model mux
     -> Qwen/Qwen3.8-Flash-Next
     -> Qwen/Qwen3.8-27B-FP8

Start with the deployment guide. Generate a random ZIPCODE_JWT_SECRET, mount a private authentication-data directory, invite the first GitHub login, and keep the origin bound to loopback behind the HTTPS edge. The inference gateway does not log request bodies. Unless disabled with ZIPCODE_DISABLE_TRACE_UPLOAD=1, the released ZIPCODE client records and uploads full agent rollouts after consent. See Trace collection and storage before onboarding a team member.

Provenance and license

ZIPCODE is derived from OpenAI Codex and retains its upstream Git history, Apache-2.0 license, and notices. ZIPCODE is an independent project and is not affiliated with, endorsed by, or sponsored by OpenAI. See NOTICE and LICENSE.