Claude-Codex Proxy Stack

Multi-model local proxy stack for Claude Code and Codex CLI — connect DeepSeek, GLM, and Qwen through protocol-translating proxies with context management, reasoning passthrough, and tool-calling compatibility. Battle-tested on Windows 11 with 17 documented edge cases resolved.

Overview

Claude Code speaks the Anthropic Messages API. Codex CLI speaks the OpenAI Responses API. Neither DeepSeek, GLM, nor Qwen implement these protocols natively. This project provides a set of local proxies that bridge the gap:

anthropic-proxy.mjs (port 7861) — unified Anthropic Messages gateway routing GLM/DeepSeek/Qwen
deepseek_proxy.py (port 9100) — Responses API to Chat Completions translator for Codex
qwen2API (port 7860, Docker) — native Anthropic endpoint for Qwen with tool-use optimization
SwitchModel — per-session model switching without touching config files
WebSearch script — search fallback for non-Anthropic models
Hook system — context monitoring, error recovery, memory persistence

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Claude Code CLI                       │
│               (Anthropic Messages API)                   │
└───────────────────────┬─────────────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────────────┐
│              anthropic-proxy.mjs :7861                   │
│         Unified context-management proxy                 │
│  (token inflation / error interception / preflight)      │
└───┬─────────────────┬─────────────────┬─────────────────┘
    │                 │                 │
┌───▼───────────┐ ┌───▼──────────┐ ┌───▼──────────────────┐
│ mydamoxing.cn │ │ api.         │ │ qwen2API :7860       │
│  GLM-5.1      │ │ deepseek.com │ │  Qwen 3.6 Plus       │
│  Anthropic    │ │ V4 Pro/Flash │ │  Native Anthropic    │
└───────────────┘ └──────────────┘ └──────────────────────┘


┌─────────────────────────────────────────────────────────┐
│                     Codex CLI                            │
│               (OpenAI Responses API)                     │
└───────────────────────┬─────────────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────────────┐
│               deepseek_proxy.py :9100                    │
│        Responses API ↔ Chat Completions translator       │
│  (reasoning passthrough / tool coercion / retry logic)   │
└───────────────────────┬─────────────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────────────┐
│                  api.deepseek.com                        │
│                  DeepSeek V4 Pro                         │
└─────────────────────────────────────────────────────────┘

Critical constraint: Clash Verge runs on 127.0.0.1:7897 and intercepts all HTTP traffic by default. All local proxies (7860/7861/9100) must be excluded via NO_PROXY=127.0.0.1,localhost,::1 — otherwise Clash returns 502 for every local request.

Features

Feature	Claude Code	Codex CLI
DeepSeek V4 Pro (1M context, reasoning)	via anthropic-proxy	via deepseek_proxy
DeepSeek V4 Flash	via anthropic-proxy	via MODEL_ALIAS
GLM-5.1 (mydamoxing)	via anthropic-proxy	—
Qwen 3.6 Plus (local Docker)	direct to qwen2API :7860/anthropic	—
Tool calling (shell, file edit, web fetch)	triple-layer defense	argument coercion + retry
Reasoning/thinking passthrough	proxy passthrough	full convert_request pipeline
Context management	token inflation + error intercept	auto-compact
Model switching per session	SwitchModel (PowerShell)	config.toml
Web search fallback	websearch.py	websearch.py
Long-term memory	MEMORY.md + interactions/	shared memory path
Boot-on-startup	—	startup folder .bat

Prerequisites

Tool	Version / Path	Purpose
Python	3.12+ (`D:/tool/py/python.exe`)	deepseek_proxy, websearch, hooks
Node.js	24+	anthropic-proxy.mjs, session-memory hook
Docker	—	qwen2API container
Git	2.50+	version control
ripgrep	15+	file search
Clash Verge	127.0.0.1:7897	network proxy (must NO_PROXY localhost)
Claude Code CLI	latest	main coding agent
Codex CLI	0.128+	secondary agent via DeepSeek

Quick Start

1. Claude Code Side

1.1 Start qwen2API (Docker)

docker run -d --name qwen2api `
  -p 7860:7860 `
  -e ADMIN_KEY=qwen2api-admin-2026 `
  -e MAX_INFLIGHT=2 `
  -e MAX_RETRIES=5 `
  yujunzhixue/qwen2api

Management console at http://127.0.0.1:7860.

1.2 Start anthropic-proxy.mjs

# Ensure NO_PROXY is set first
$env:NO_PROXY = "127.0.0.1,localhost,::1"
$env:HTTP_PROXY = ""

node ~/Desktop/qwen2api/anthropic-proxy.mjs

This proxy listens on port 7861 and routes requests based on ANTHROPIC_MODEL:

glm-5.1 / glm-4.7 → mydamoxing.cn (Anthropic passthrough)
deepseek-v4-pro / deepseek-v4-flash → api.deepseek.com (Anthropic passthrough)
qwen3.6-plus → qwen2API :7860 (Anthropic→OpenAI translation)

1.3 Configure Claude Code

Create ~/.claude/glm-settings.json (or equivalent for your default model):

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:7861",
    "ANTHROPIC_MODEL": "glm-5.1",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.1",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.1",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.1",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "32768",
    "MAX_MCP_OUTPUT_TOKENS": "32768"
  },
  "modelOverrides": {
    "claude-4.5-opus": "glm-4.7",
    "claude-4.6-sonnet": "glm-4.7",
    "claude-4.5-haiku": "glm-4.7"
  },
  "maxTokens": 32768,
  "autoCompactThreshold": 0.7
}

Launch with:

claude --settings ~/.claude/glm-settings.json

1.4 SwitchModel (Per-Session Model Switching)

SwitchModel is a PowerShell 7 function (sourced from ~/.claude/scripts/switch.ps1). It sets process-level environment variables only — no file writes, auto-cleanup on terminal close.

SwitchModel glm            # GLM 5.1 (mydamoxing)
SwitchModel deepseek-pro   # DeepSeek V4 Pro
SwitchModel deepseek-flash # DeepSeek V4 Flash
SwitchModel qwen           # Qwen 3.6 Plus (local :7860/anthropic)
SwitchModel claude         # Clear override — back to settings.json default
SwitchModel status         # Show current session model

2. Codex CLI Side

2.1 Start deepseek_proxy.py

# start_proxy.bat — double-click or run in terminal
@echo off
netstat -ano | findstr ":9100" | findstr LISTENING >nul 2>&1
if %ERRORLEVEL% EQU 0 exit /b 0
set HTTP_PROXY=
set HTTPS_PROXY=
start "" /B "D:\tool\py\pythonw.exe" "C:\Users\<user>\.codex\deepseek_proxy.py" 9100

This launches Flask + waitress on port 9100 in background (no console window). All stdout/stderr redirect to ~/.codex/proxy.log.

2.2 Configure Codex

~/.codex/config.toml:

model = "gpt-5.5"
model_provider = "deepseek"
model_reasoning_effort = "high"
model_reasoning_summary = "auto"
ask_for_approval = "never"
sandbox = "danger-full-access"

[model_providers.deepseek]
name = "DeepSeek"
base_url = "http://127.0.0.1:9100/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

Why model = "gpt-5.5"? Codex has built-in metadata for GPT-5.5 — it correctly sends reasoning.effort in requests and uses shell_command (string format). The proxy maps it back to deepseek-v4-pro before forwarding upstream.

2.3 Launch Codex

# start_codex.bat
@echo off
chcp 65001 >nul
call "%USERPROFILE%\.codex\start_proxy.bat"
set HTTP_PROXY=
set HTTPS_PROXY=
set NO_PROXY=127.0.0.1,localhost,::1
set no_proxy=127.0.0.1,localhost,::1
codex %*

Component Details

anthropic-proxy.mjs (Port 7861)

Location: ~/Desktop/qwen2api/anthropic-proxy.mjs

The unified context-management proxy. All three models pass through port 7861 to reach Claude Code.

Key capabilities:

Model routing — dispatches by ANTHROPIC_MODEL env var to the correct upstream
Triple-layer context defense:
1. count_tokens inflation — when tokens exceed 80% of model window, reports 180K~200K to trigger Claude Code auto-compact
2. Error interception — translates upstream context-limit errors to Anthropic standard format
3. Request preflight — rejects requests exceeding model window with HTTP 400
Tool-calling passthrough — preserves tool schemas for DeepSeek/GLM

API keys configured in-code:

GLM (mydamoxing):
DeepSeek:
Qwen:

deepseek_proxy.py (Port 9100)

Location: ~/.codex/deepseek_proxy.py (~520 lines)

A Flask + waitress server that translates OpenAI Responses API (Codex native) to DeepSeek Chat Completions API.

Key capabilities:

convert_request(data) — full protocol translation:
- Buffers reasoning items and attaches reasoning_content to the next assistant message (required by DeepSeek thinking mode)
- Merges parallel function_call items into a single assistant message (avoids DeepSeek 400 error)
- Maps gpt-5.5 → deepseek-v4-pro via MODEL_ALIAS
- Sets thinking.enabled for pro models, thinking.disabled for flash or minimal effort
coerce_tool_args(name, args_str) — bidirectional argument conversion:
- shell_command: array ["bash","-lc","cmd"] → string "cmd" (GPT-5.5 requirement)
- apply_patch: JSON → raw patch text
- Legacy shell/local_shell: string → array
fetch_chunks() — dual-mode upstream:
- Tools present → requests library, non-streaming, 3 retries (1/2/3s backoff)
- No tools → httpx library, true streaming, auto-fallback to non-streaming on disconnect
SSE state machine — lazily creates reasoning/message/tool_call items, ensures correct event ordering (reasoning closed before message opened), all done events carry accumulated text
pythonw compatibility — detects sys.stdout is None and redirects to proxy.log

Debug files:

~/.codex/last_inbound.json — raw request from Codex
~/.codex/last_request.json — converted request sent to DeepSeek
~/.codex/proxy.log — runtime stdout/stderr

qwen2API (Port 7860, Docker)

Image: yujunzhixue/qwen2api

Exposes three API formats on port 7860:

/v1/chat/completions — OpenAI format
/anthropic/v1/messages — Native Anthropic with tool-calling optimization (schema compression, tool name obfuscation, few-shot injection)
/v1beta/models/... — Gemini format

Important: Claude Code must connect to /anthropic/v1/messages directly. Routing through anthropic-proxy.mjs (which does Anthropic→OpenAI translation) bypasses qwen2API's optimization code, resulting in 0-token responses and no tool usage.

Docker configuration:

MAX_INFLIGHT=2
MAX_RETRIES=5
ACCOUNT_MIN_INTERVAL_MS=600
REQUEST_JITTER=120~360ms
shm_size=512m

SwitchModel

Location: ~/.claude/scripts/switch.ps1, registered as a PowerShell 7 function.

Per-session model switching. Sets process-level $env:ANTHROPIC_* variables only — never writes to config files. Terminal close = auto-cleanup.

WebSearch Script

Location: ~/.claude/scripts/websearch.py

Claude Code's native WebSearch tool only works with official Anthropic models. For GLM/DeepSeek/Qwen, use this script instead:

D:/tool/py/python.exe ~/.claude/scripts/websearch.py "query"
D:/tool/py/python.exe ~/.claude/scripts/websearch.py -n 10 "query"
D:/tool/py/python.exe ~/.claude/scripts/websearch.py --news "query"
D:/tool/py/python.exe ~/.claude/scripts/websearch.py --json "query"

Supports -n (result count), --news (time-sensitive), --json (structured output), and --backend (search engine selection).

Hook System (Claude Code)

Hook	Trigger	Purpose
`pre-compact-hook.py`	PreCompact	Update STATE.md before context compression
`session-start-hook.py`	SessionStart	Inject CWD/STATE.md + memory index
`context-monitor-hook.py`	UserPromptSubmit	Check context water level each input
`error-recovery-hook.py`	Notification (error)	Trigger 6-step self-repair instructions
`session-memory-hook.cjs`	SessionStart (compact/clear)	Record session summary
`observation-logger.cjs`	PostToolUse	Log key observations after each tool call
`statusline-wrapper.js`	statusLine	Correct context window percentage display

Memory System

Location: ~/.claude/projects/C--Users----/memory/

A markdown-based long-term memory system indexed by MEMORY.md. Categories:

user_*.md — user profile
feedback_*.md — lessons learned (42 files)
project_*.md — project states
reference_*.md — configuration references
interactions/ — dynamic data (quiz records, study progress)

AGENTS.md / CLAUDE.md

Global agent instructions injected into system context:

~/.codex/AGENTS.md — Codex-specific: shell priority (exe > bash > powershell), GPT-5.5 string-format commands, websearch fallback
~/.claude/CLAUDE.md — Claude-specific: zero-attribution policy, coding principles, error handling, STATE.md handoff

Known Issues & Fixes

Clash Intercepts Localhost (Issue #1)

Symptom: curl 127.0.0.1:9100 returns 502. Diagnosis: curl -v shows Trying 127.0.0.1:7897. Fix: Set NO_PROXY=127.0.0.1,localhost,::1 in 4 places: setx (system-wide), bashrc, PowerShell profile, startup scripts. Also clear HTTP_PROXY.

Codex 0.128 Requires wire_api=responses (Issue #2)

Symptom: wire_api=chat causes startup error. Fix: Must use wire_api=responses. The local proxy is the only path.

httpx Streaming + DeepSeek Multi-byte Chars (Issue #3)

Symptom: peer closed connection without sending complete message body (RemoteProtocolError). Root cause: httpx 0.28 chunked tail decode fails with multi-byte characters. Fix: Switch to requests non-streaming when tools are present. Auto-fallback on disconnect without tools.

DeepSeek Stream Drops on Tool Args (Issue #4)

Symptom: Upstream server disconnects mid-stream when tool arguments are present. Fix: Force non-streaming when tools are in the request.

Random ChunkedEncodingError (Issue #5)

Symptom: Intermittent upstream errors. Fix: 3 retries with 1/2/3-second backoff in proxy.

Codex Doesn't Send reasoning.effort for Non-OpenAI Providers (Issue #6)

Symptom: model_reasoning_effort=high in config but absent from request. Fix: Masquerade as gpt-5.5 (Codex has built-in metadata). Proxy defaults thinking=enabled for pro models.

pythonw stdout=None Crashes Flask (Issue #7)

Symptom: Background pythonw process crashes on first print. Fix: Detect sys.stdout is None at proxy start and redirect to proxy.log.

Codex Forces PowerShell by Default (Issue #8)

Symptom: Model uses PowerShell for every command — slow startup, awkward Unix commands. Fix: AGENTS.md explicitly states shell priority: direct exe > bash > powershell. Empirically overrides default behavior.

Streaming Done Events Had Empty Text (Issue #9)

Symptom: Codex TUI shows blank output but token count increments. Fix: response.content_part.done and response.output_item.done must carry accumulated text.

DeepSeek Thinking Mode Requires Historical reasoning_content (Issue #10)

Symptom: The reasoning_content in the thinking mode must be passed back to the API. Fix: convert_request buffers reasoning item summary text and attaches it as reasoning_content to the next assistant message.

Streaming State Machine Indentation Bug (Issue #11)

Symptom: Stream output blank because all chunk processing was nested inside if not response_started:. Fix: Moved the condition inside the for-loop at correct indentation level.

shell_command Argument Format Mismatch (Issue #12)

Symptom: Model outputs array ["bash","-lc","..."], GPT-5.5 expects string → failed to parse: expected a string. Fix: coerce_tool_args handles shell_command array→string conversion.

Parallel Tool Call Splitting Causes DeepSeek 400 (Issue #13)

Symptom: 8 parallel function_calls become 8 separate assistant messages → DeepSeek returns insufficient tool messages → Codex treats empty response as task complete. Fix: Buffer consecutive function_call items, flush as a single merged assistant message.

Context Window Percentage Display (Issue #14)

Symptom: Status bar shows inflated context percentage. Root cause: CLI calculates denominator as 200K for proxy models; actual window is 1M. Fix: statusline-wrapper.js rewrites context_window_size based on model name.

CLAUDE_CODE_CONTEXT_WINDOW is Not a Valid Env Var (Issue #15)

Fix: Deleted the invalid env var. Use statusline-wrapper.js for display fix. Rely on autoCompactThreshold for actual compaction.

mydamoxing glm-5.1 Only Has OpenAI Endpoint (Issue #16)

Symptom: Direct Anthropic endpoint connection fails for glm-5.1. Fix: Route through 7861 proxy which does Anthropic→OpenAI translation for glm-5.1.

qwen2API Must Use Native Anthropic Endpoint (Issue #17)

Symptom: Routing through 7861 translation layer causes Qwen to return 0 tokens, no tool calls. Root cause: Bypasses qwen2API's built-in Anthropic optimization (schema compression, tool name obfuscation, few-shot injection). Fix: Claude Code connects directly to :7860/anthropic/v1/messages.

Configuration Reference

Required Environment Variables

# System-wide (setx)
NO_PROXY=127.0.0.1,localhost,::1

# Claude Code (in settings.json env or SwitchModel)
ANTHROPIC_BASE_URL=http://127.0.0.1:7861
ANTHROPIC_AUTH_TOKEN=<your-key>
ANTHROPIC_MODEL=glm-5.1
CLAUDE_CODE_MAX_OUTPUT_TOKENS=32768

# Codex (in config.env)
OPENAI_API_KEY=<placeholder>
DEEPSEEK_API_KEY=<your-key>
NO_PROXY=127.0.0.1,localhost

File Map

~/
├── .claude/
│   ├── CLAUDE.md                    # Global agent rules
│   ├── settings.json                # Claude Code config
│   ├── glm-settings.json            # GLM default model config
│   ├── scripts/
│   │   ├── switch.ps1               # SwitchModel function
│   │   ├── websearch.py             # Web search fallback
│   │   ├── pre-compact-hook.py      # STATE.md update before compact
│   │   ├── session-start-hook.py    # Memory index injection
│   │   ├── context-monitor-hook.py  # Context water level check
│   │   ├── error-recovery-hook.py   # Error-triggered repair
│   │   └── statusline-wrapper.js    # Context percentage fix
│   ├── session-memory/
│   │   ├── session-memory-hook.cjs  # Session summary logger
│   │   └── observation-logger.cjs   # Tool observation logger
│   └── projects/C--Users----/memory/
│       ├── MEMORY.md                # Memory index
│       └── ...                      # Memory files
├── .codex/
│   ├── config.toml                  # Codex model/provider config
│   ├── config.env                   # API keys + NO_PROXY
│   ├── AGENTS.md                    # Global agent rules
│   ├── deepseek_proxy.py            # Responses→Chat Completions proxy
│   ├── start_proxy.bat              # Launch proxy (pythonw)
│   ├── start_codex.bat              # Launch Codex with proxy
│   ├── last_inbound.json            # Debug: raw Codex request
│   ├── last_request.json            # Debug: converted request
│   └── proxy.log                    # Debug: proxy stdout/stderr
└── Desktop/qwen2api/
    ├── anthropic-proxy.mjs          # Unified context proxy
    └── start_proxy.bat              # Launch with NO_PROXY

Debugging

Codex + DeepSeek

Check	Command
Is proxy running?	`netstat -ano
Test proxy directly	`curl --noproxy 127.0.0.1 http://127.0.0.1:9100/v1/models`
Test DeepSeek directly	`curl https://api.deepseek.com/v1/models -H "Authorization: Bearer <key>"`
View raw Codex request	`cat ~/.codex/last_inbound.json`
View converted request	`cat ~/.codex/last_request.json`
View proxy logs	`cat ~/.codex/proxy.log`
Restart proxy	`taskkill /F /PID <pid>` then `cmd /c ~/.codex/start_proxy.bat`

Claude + Multi-Model

Check	Command
Memory index	`cat ~/.claude/projects/C--Users----/memory/MEMORY.md`
qwen2API console	`http://127.0.0.1:7860`
Current model	`SwitchModel status`
Global config	`cat ~/.claude/glm-settings.json`

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
README_CN.md		README_CN.md

Folders and files

Latest commit

History

Repository files navigation