Claude Code Gateway Workflow on a Public Research Snapshot

Run Claude-style CLI workflows through a local LiteLLM gateway with OpenAI, Gemini, DeepSeek, or OpenRouter.

This repo is unofficial, local-first, and layered onto a public Claude Code source snapshot. Upstream ownership remains with Anthropic.

What This Repo Is Useful For

• Trying a Claude-style terminal workflow against a local Anthropic-compatible gateway
• Switching providers without rewriting the CLI itself
• Rebuilding and inspecting the Bun + Ink CLI from source
• Studying the exposed snapshot and its architecture for defensive research

Quick Start

1. Install the local dependencies:

   bun install
   python3 -m pip install "litellm[proxy]"
   

2. Create a safe local env file and check your machine:

   ./bootstrap-gateway.sh
   

3. Open examples/provider-gateway/gateway.env and fill in one provider key:

   • OPENAI_API_KEY
   • GOOGLE_API_KEY or GEMINI_API_KEY
   • DEEPSEEK_API_KEY
   • OPENROUTER_API_KEY

4. Pick a route:

   ./use-provider auto
   # or
   ./use-provider openai openai-neon "My Local CLI"
   # or
   ./use-provider openrouter
   

5. Validate the setup:

   ./doctor-gateway
   ./scripts/smoke-test-gateway.sh
   

6. Start LiteLLM in one terminal:

   cd examples/provider-gateway
   ./start-litellm.sh
   

7. Start the CLI from the repo root in another terminal:

   ./run-claude.sh
   

examples/provider-gateway/gateway.env is local-only and gitignored. Do not commit real API keys.

What Works / What Does Not

Area	Current status	Notes
Bun build from source	Supported in this fork	bun install && bun run build
Local LiteLLM gateway routing	Supported in this fork	Routes included for OpenAI, Gemini, DeepSeek, and OpenRouter
Provider switching helpers	Supported in this fork	./use-provider and ./use-openai update the local gateway env
Local diagnostics	Supported in this fork	./doctor-gateway and ./scripts/smoke-test-gateway.sh
Official Anthropic support	Not included	This repo is unofficial and not maintained by Anthropic
Exact Anthropic parity on non-Anthropic backends	Not guaranteed	Some models or side features may behave differently through LiteLLM
Hosted gateway or managed secrets	Not included	The workflow is designed for local development
Clean-room reimplementation or upstream relicensing	Not included	This remains a public research/archive snapshot with fork-added tooling

Gateway Mode

In gateway mode, the CLI still speaks Anthropic-shaped HTTP, but it does so against a local LiteLLM server instead of Anthropic's hosted API.

Flow:

1. The CLI reads examples/provider-gateway/gateway.env.
2. ANTHROPIC_BASE_URL points the CLI at local LiteLLM, usually http://127.0.0.1:4000.
3. ANTHROPIC_MODEL selects one of the route names defined in examples/provider-gateway/litellm.yaml.
4. LiteLLM forwards the request to the provider whose API key you configured.

Default route names in this repo:

• openai-via-gateway
• gemini-via-gateway
• deepseek-via-gateway
• openrouter-via-gateway

The bundled litellm.yaml uses drop_params: true for better cross-provider compatibility.

Exact Commands

Build the CLI:

bun install
bun run typecheck
bun run build

Bootstrap the local env:

./bootstrap-gateway.sh

Pick a provider:

./use-provider openrouter
./use-provider openai openai-neon
./use-provider gemini cyber-fox "Gemini Gateway CLI"
./use-provider auto
./use-openai

Validate before launch:

./doctor-gateway
./scripts/smoke-test-gateway.sh

Start LiteLLM:

cd examples/provider-gateway
./start-litellm.sh

Run the CLI:

cd /path/to/claude-code-main
./run-claude.sh

Helper Scripts

Script	What it does	Typical use
./bootstrap-gateway.sh	Confirms repo context, checks Bun/Python/LiteLLM, and creates a safe local gateway.env if missing	First run on a new machine
./use-provider	Switches ANTHROPIC_MODEL, ANTHROPIC_SMALL_FAST_MODEL, and optional branding	Swap between OpenAI, Gemini, DeepSeek, and OpenRouter
./use-openai	Shortcut wrapper for ./use-provider openai	Faster OpenAI selection
./doctor-gateway	Verifies local env, route names, provider keys, and branding values	Quick readiness check before launch
./run-claude.sh	Root-level wrapper that launches the example gateway workflow	Normal local CLI entrypoint
examples/provider-gateway/start-litellm.sh	Starts the local LiteLLM proxy using the bundled config	Terminal 1
examples/provider-gateway/run-claude.sh	Builds the CLI if needed, then runs it with the local gateway env	Terminal 2
./scripts/smoke-test-gateway.sh	Runs safe local validation without making real provider calls	CI and local sanity checks

Safety / Secrets

• examples/provider-gateway/gateway.env is for local use only and is already gitignored.
• examples/provider-gateway/gateway.env.example is the safe template meant for the repository.
• In gateway mode, ANTHROPIC_API_KEY is the local LiteLLM master key expected by litellm.yaml, not an Anthropic cloud credential.
• Provider keys should stay only in your local gateway.env or your shell environment.
• Before pushing changes, double-check that no .env file or real API key is staged.

Troubleshooting

Symptom	Likely cause	Fix
Missing .../gateway.env	Local env file has not been created yet	Run ./bootstrap-gateway.sh
Selected ... but ... API KEY is empty	The chosen provider route does not have a matching key set	Edit examples/provider-gateway/gateway.env and add the matching key
model route not found in litellm.yaml	ANTHROPIC_MODEL does not match a configured route name	Run ./use-provider ... again or update litellm.yaml deliberately
LiteLLM is not on PATH	LiteLLM is not installed in your current Python environment	Run python3 -m pip install "litellm[proxy]"
bun not found	Bun is not installed or your shell has not reloaded PATH	Install Bun, reopen the terminal, and rerun ./bootstrap-gateway.sh
dist/cli.js missing	The CLI has not been built yet	Run bun run build or just launch ./run-claude.sh and let it build
Smoke test fails in CI for provider keys	The example env intentionally leaves provider keys empty	Run the smoke test with --ci or use a local env file with a real provider key

Practical Setup Doc

For a shorter, task-oriented walkthrough, see docs/PRACTICAL_SETUP.md.

Unofficial Status and Ownership

This repository contains a mirrored Claude Code source snapshot plus fork-added gateway tooling.

• The original Claude Code source and related intellectual property are owned by Anthropic.
• This repository is not affiliated with, endorsed by, or maintained by Anthropic.
• Except where explicitly stated, no rights are granted to use, reproduce, sublicense, or redistribute upstream Anthropic source code.
• This repository should not be presented as official Anthropic software.

Fork-Added Files and Licensing Scope

Only files authored in this fork are licensed under LICENSE.fork-additions (MIT), and only for portions authored by the fork maintainer.

Fork-added or fork-authored files include:

• doctor-gateway
• use-provider
• use-openai
• run-claude.sh (root wrapper)
• bootstrap-gateway.sh
• scripts/smoke-test-gateway.sh
• examples/provider-gateway/start-litellm.sh
• examples/provider-gateway/run-claude.sh
• examples/provider-gateway/SETUP.txt
• examples/provider-gateway/litellm.yaml
• examples/provider-gateway/gateway.env.example
• BUILD_GUIDE.md
• docs/PRACTICAL_SETUP.md
• README sections explicitly describing fork-added tooling and workflows

All other files remain subject to their original copyright and license status (if any), and are excluded from the fork-added MIT grant.

Research / Archive Context

This repository still exists as a defensive research and source-exposure archive.

• It mirrors a publicly exposed Claude Code source snapshot that became accessible on March 31, 2026 through an npm source-map exposure.
• It is maintained for education, defensive security research, and software supply-chain analysis.
• It is not a clean-room rewrite and not a claim of upstream ownership.

Related research writing:

• Original publication — Hong Minhee, Is legal the same as legitimate: AI reimplementation and the erosion of copyleft

The essay is dated March 9, 2026 and serves as companion analysis that predates the March 31, 2026 source exposure documented in this repo.

Why This Archive Exists

The snapshot is useful for:

• educational study
• software supply-chain exposure analysis
• secure software engineering discussion
• architecture review of modern agentic CLI systems

How the Public Snapshot Became Accessible

Chaofan Shou (@Fried_rice) publicly noted that Claude Code source material was reachable through an exposed .map file in the npm package, which in turn referenced unobfuscated TypeScript sources hosted in Anthropic's R2 bucket.

Repository Scope

Claude Code is Anthropic's CLI for interacting with Claude from the terminal to perform engineering tasks such as editing files, running commands, searching codebases, and coordinating workflows.

This repository contains a mirrored src/ snapshot for research and analysis, plus fork-added local gateway tooling.

• Language: TypeScript
• Runtime: Bun
• Terminal UI: React + Ink
• Extra local workflow: LiteLLM provider gateway under examples/provider-gateway/

Built with oh-my-codex

The README and archive-context work on this branch were AI-assisted and orchestrated with Yeachan Heo's oh-my-codex (OmX), a Codex workflow layer.

• $team mode was used for coordinated review of repo fit, wording risk, and architecture consistency.
• $ralph mode was used for persistent execution, verification, and final sign-off before completion.
• This repository does not claim ownership of the original code and should not be interpreted as an official Anthropic repository.

Architecture Notes

High-Level Layout

src/
├── main.tsx                 # CLI startup and initialization
├── commands.ts              # Command registry
├── tools.ts                 # Tool registry
├── QueryEngine.ts           # LLM query engine
├── commands/                # Slash command implementations
├── tools/                   # Tool implementations
├── components/              # Ink UI components
├── services/                # Service integrations
├── bridge/                  # IDE bridge system
├── coordinator/             # Multi-agent orchestration
├── plugins/                 # Plugin loader
└── skills/                  # Skill system

Key Subsystems

Subsystem	What it does
src/tools/	Tool definitions, schemas, permissions, and execution logic
src/commands/	Slash command implementations such as review, config, tasks, and doctor flows
src/services/	API, MCP, OAuth, LSP, and integration services
src/bridge/	IDE bridge code for editors such as VS Code and JetBrains
src/coordinator/	Multi-agent coordination and team workflows
src/plugins/	Built-in and third-party plugin support
src/skills/	Reusable workflow definitions executed through the skill system

Notable Design Patterns

• Parallel startup prefetch for settings, keychain reads, and network setup
• Lazy loading for heavier integrations and feature-gated systems
• Tool-driven orchestration rather than a single monolithic agent loop
• Bridge-based editor integration for IDE-connected workflows