@codeproxy/cli

中文版 → README.zh-CN.md

@codeproxy/cli is a local proxy server that converts any Chat Completions or Anthropic Messages API into the OpenAI Responses API format. It lets Codex, Claude Code, or any Responses-API client use models from DeepSeek, GLM, Kimi, and more.

Built on @codeproxy/core.

Quick Start

npx @codeproxy/cli --upstream-format openai-chat \
  --base-url https://api.deepseek.com/v1 \
  --apikey sk-your-key

Point your Responses-API client at http://127.0.0.1:8787:

curl -N http://127.0.0.1:8787/v1/responses \
  -H 'content-type: application/json' \
  -H "authorization: Bearer \$API_KEY" \
  -d '{"model":"deepseek-v4-pro","input":"Hello!","stream":true}'

With config file

npx @codeproxy/cli --config config.json

See config.example.json for a full example.

Top-level fields

Field	Type	Description
`version`	`string`	Config format version (currently `"1.0"`)
`currentUpstream`	`string`	Name of the upstream to use (must match a key in `upstreams`)
`headers`	`object`	Default headers applied to all upstreams (merged with per-upstream headers; per-upstream wins)
`reasoningEffort`	`string`	Default reasoning effort for all upstreams: `"low"`, `"medium"`, `"high"`, `"xhigh"`
`thinking`	`object` `	null`
`timeoutMs`	`number`	Default upstream request timeout in milliseconds

Per-upstream fields

Field	Type	Description
`format`	`"anthropic"` `	"openai-chat"`
`baseUrl`	`string`	Required. Upstream endpoint URL
`apiKey`	`string`	Upstream API key. Sent as `Authorization: Bearer <key>` (Anthropic: rewritten to `x-api-key`)
`model`	`string`	Override the `model` field in all incoming requests
`apiVersion`	`string`	Override `anthropic-version` header (Anthropic only)
`headers`	`object`	Extra HTTP headers for this upstream. Merged on top of top-level `headers`, per-upstream wins
`timeoutMs`	`number`	Request timeout for this upstream (overrides top-level `timeoutMs`)
`dropImages`	`boolean`	When `true`, strip image/file parts from user messages (for text-only models). Use with `fallback` to auto-route image requests to a vision-capable upstream
`fallback`	`string`	Name of another upstream to route to when `dropImages: true` and the request contains images
`reasoningEffort`	`string`	Per-upstream reasoning effort override (`"low"`, `"medium"`, `"high"`, `"xhigh"`). Overrides top-level value
`thinking`	`object` `	null`

Precedence

CLI flags > per-upstream fields > top-level fields > built-in defaults

Example: auto-fallback for text-only models

When deepseek has dropImages: true and the user sends an image, the proxy automatically routes to kimi-vision (uses Kimi K2.6):

{
  "currentUpstream": "deepseek",
  "upstreams": {
    "deepseek": {
      "baseUrl": "https://api.deepseek.com/v1",
      "apiKey": "sk-...",
      "model": "deepseek-v4-pro",
      "dropImages": true,
      "fallback": "kimi-vision"
    },
    "kimi-vision": {
      "baseUrl": "https://api.moonshot.cn/v1",
      "apiKey": "sk-...",
      "model": "kimi-k2.6",
      "headers": { "x-llm-api-key": "sk-..." }
    }
  }
}

Codex Configuration

Codex 0.128.0+ requires custom providers to speak the Responses API. @codeproxy/cli bridges this gap for any Chat Completions or Anthropic Messages upstream.

Quick setup

Start the proxy:

npx @codeproxy/cli --upstream-format openai-chat \
  --base-url https://api.deepseek.com/v1 \
  --apikey sk-your-key

Add a custom provider in ~/.codex/config.toml:

[model_providers.deepseek]
name = "DeepSeek"
base_url = "http://127.0.0.1:8787/v1"
wire_api = "responses"

[profiles.deepseek-pro]
model = "deepseek-v4-pro"
model_provider = "deepseek"

With reasoning effort

[model_providers.deepseek]
name = "DeepSeek"
base_url = "http://127.0.0.1:8787/v1"
wire_api = "responses"

[profiles.deepseek-pro]
model = "deepseek-v4-pro"
model_provider = "deepseek"
model_reasoning_effort = "high"

@codeproxy/cli automatically maps reasoning.effort to the upstream's native format (e.g., reasoning_effort for OpenAI Chat, thinking blocks for Anthropic).

Multiple upstreams via config file

{
  "currentUpstream": "deepseek-chat",
  "upstreams": {
    "deepseek-chat": {
      "baseUrl": "https://api.deepseek.com/v1",
      "apiKey": "sk-...",
      "model": "deepseek-v4-pro",
      "dropImages": true,
      "fallback": "kimi-vision"
    },
    "kimi-vision": {
      "baseUrl": "https://api.moonshot.cn/v1",
      "apiKey": "sk-...",
      "model": "kimi-k2.6",
      "headers": { "x-llm-api-key": "sk-..." }
    },
    "claude": {
      "format": "anthropic",
      "baseUrl": "https://api.anthropic.com/v1",
      "apiKey": "sk-ant-...",
      "model": "claude-sonnet-4-20250514"
    }
  }
}

Switch upstreams by changing currentUpstream and restarting the proxy — no Codex config changes needed.

Install

npm install -g @codeproxy/cli

CLI Options

Flag	Default	Description
`--base-url <url>`	—	Upstream endpoint (required unless `--config`)
`--upstream-format <fmt>`	inferred	`anthropic` or `openai-chat`
`--config <file>`	—	JSON config file
`--host <host>`	`127.0.0.1`	Bind host
`-p, --port <port>`	`8787`	Bind port
`--api-version <ver>`	`2023-06-01`	Override Anthropic version header
`--apikey <key>`	—	Upstream API key
`--model <name>`	—	Override model for all requests
`--drop-images`	—	Strip images (text-only models)

Programmatic usage

import { createResponsesFetch, startProxy } from '@codeproxy/cli';

const proxy = await startProxy({
  upstreamFormat: 'openai-chat',
  baseUrl: 'https://api.deepseek.com/v1',
  defaultHeaders: { authorization: 'Bearer sk-...' },
});

License

MIT