p5js.ai 2 API

p5js.ai 2 API — 将 p5js.ai 免费接口包装为 Anthropic / OpenAI 兼容 API

p5js.ai 2 API 是一个轻量级的反向代理适配器，将 https://p5js.ai/api/ai-chat 免费聊天接口同时包装成 Anthropic Messages API 和 OpenAI Chat Completions API 兼容接口。

这意味着任何支持 Anthropic 或 OpenAI 协议的工具——Claude Code、Chatbox、NextChat、LobeChat、one-api、Cherry Studio 等——都可以直接接入，就像连接一个真正的 Anthropic 或 OpenAI 端点一样。

核心特性

双协议兼容 — 同时提供 /v1/messages（Anthropic）和 /v1/chat/completions（OpenAI）端点
Tool Use 仿真 — 上游不支持原生 tool_use，本服务通过 XML 格式的 <function_calls> 提示词注入实现伪工具调用，并自动将响应中的 XML 解析回标准 tool_use / tool_calls 格式
p5.js 噪声过滤 — 自动检测并剥离上游注入的 p5.js 助手问候语、前缀和尾缀
SSE 修复 — 上游会输出畸形的 ddata: / ata: 前缀，本服务自动修正为标准 SSE 格式
双层缓存 — 内存缓存 + 可选 Redis 二级缓存，相同请求自动命中，流式/非流式共用同一份完成结果
In-flight 合并 — 并发相同请求不会重复打上游，follower 等待 leader 结果
连接池复用 — 共享 httpx 异步连接池，适合高并发场景
代理支持 — 支持 UPSTREAM_PROXY_URL 或标准 HTTP_PROXY / HTTPS_PROXY 环境变量

项目结构

p5js/
├── main.py              # FastAPI 应用入口、路由定义
├── config.py            # 常量、环境变量、模型列表、正则模式
├── filters.py           # p5.js 噪声过滤、工具感知文本缓冲
├── tools.py             # Tool XML 提示词构建、解析、提取
├── translate.py         # 协议转换（Anthropic/OpenAI → 上游消息格式）
├── upstream.py          # 上游 HTTP 客户端、SSE 解析、实时捕获
├── render.py            # Artifact → Anthropic/OpenAI JSON/SSE 渲染
├── stream.py            # 实时流处理（带/不带工具，带缓存集成）
├── response_cache.py    # 双层缓存系统（内存 + Redis）
├── tests/               # 测试套件
│   ├── test_response_cache.py
│   └── test_filters.py
├── Dockerfile
├── docker-compose.yml
├── requirements.txt
├── start.sh
├── .env.example         # 环境变量示例
└── README.md

快速开始

本地运行

./start.sh

首次运行会自动创建 venv 并安装依赖，然后在 http://127.0.0.1:18185 监听。

Docker 部署

# 直接构建
docker build -t p5js2api:latest .
docker run --rm -p 18185:18185 p5js2api:latest

# 或使用 docker compose（自带 Redis）
docker compose up -d --build

默认 compose 配置同时启动一个本地 Redis 实例，端口通过 P5JS2API_PORT 环境变量控制（默认 18185）。

支持的模型

模型	说明
`claude-opus-4-7`	最新旗舰
`claude-opus-4-6`
`claude-opus-4-1` / `claude-opus-4-1-20250805`
`claude-opus-4-20250514`
`claude-sonnet-4-6`
`claude-sonnet-4-5` / `claude-sonnet-4-5-20250929`	默认模型
`claude-sonnet-4-20250514`
`claude-haiku-4-5` / `claude-haiku-4-5-20251001`	轻量快速

接口一览

路径	方法	说明
`/health`	GET	健康检查，返回服务状态、缓存统计、上游配置
`/v1/models`	GET	模型列表（Anthropic 格式）
`/v1/messages`	POST	Anthropic Messages API，支持 `stream`
`/v1/chat/completions`	POST	OpenAI Chat Completions API，支持 `stream`

服务不校验 API Key，任意非空字符串均可通过认证。

使用示例

Claude Code（Anthropic 协议）

export ANTHROPIC_BASE_URL=http://127.0.0.1:18185
export ANTHROPIC_AUTH_TOKEN=dummy
export ANTHROPIC_MODEL=claude-opus-4-7
export ANTHROPIC_SMALL_FAST_MODEL=claude-haiku-4-5

claude

OpenAI Python SDK

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:18185/v1",
    api_key="sk-dummy",
)

resp = client.chat.completions.create(
    model="claude-sonnet-4-5",
    messages=[{"role": "user", "content": "你好"}],
    stream=True,
)
for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="", flush=True)

第三方工具（Chatbox / NextChat / LobeChat / one-api / Cherry Studio 等）

Base URL / API 地址: http://127.0.0.1:18185/v1
API Key: 任意非空字符串（如 sk-dummy）
模型名: 填写上方「支持的模型」中的任一项

curl

# Anthropic 协议
curl http://127.0.0.1:18185/v1/messages \
  -H 'content-type: application/json' \
  -H 'x-api-key: dummy' \
  -d '{"model":"claude-sonnet-4-5","max_tokens":1024,"messages":[{"role":"user","content":"hi"}]}'

# OpenAI 协议
curl http://127.0.0.1:18185/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"claude-sonnet-4-5","messages":[{"role":"user","content":"hi"}]}'

缓存系统

服务默认启用完整响应缓存：

先走进程内内存缓存（LRU，默认 256 条）
配置 RESPONSE_CACHE_REDIS_URL 后升级为 内存 + Redis 双层缓存
相同请求并发命中 miss 时做 in-flight 合并，避免同时打爆上游
stream=true 和 stream=false 共用同一份完成结果缓存

环境变量

完整环境变量列表见 .env.example，核心配置：

变量	默认值	说明
`RESPONSE_CACHE_ENABLED`	`true`	是否启用缓存
`RESPONSE_CACHE_TTL_SECS`	`300`	普通请求缓存 TTL（秒）
`RESPONSE_CACHE_TOOL_TTL_SECS`	`120`	带工具请求缓存 TTL（秒）
`RESPONSE_CACHE_MAX_ENTRY_BYTES`	`33554432`	单条缓存最大字节数（0 = 不限）
`RESPONSE_CACHE_REDIS_URL`	—	Redis 连接地址，配置后启用二级缓存

响应头

X-Proxy-Cache: HIT | MISS | BYPASS | DISABLED
X-Proxy-Cache-Source: memory | redis | inflight | live

跳过缓存

任一方式：

请求头 X-Proxy-Cache: bypass
请求头 Cache-Control: no-cache

代理配置

容器内支持两种代理方式：

显式设置 UPSTREAM_PROXY_URL
标准环境变量 HTTP_PROXY / HTTPS_PROXY / NO_PROXY

/health 里的 upstream.proxy_configured 会显示当前是否检测到代理配置。

工作原理

┌─────────────┐     ┌──────────────────┐     ┌─────────────┐
│  Client      │────▶│  p5js.ai 2 API   │────▶│  p5js.ai    │
│ (Claude/     │◀────│  (this project)   │◀────│  upstream   │
│  OpenAI)     │     │                   │     │             │
└─────────────┘     └──────────────────┘     └─────────────┘
                    │  ├─ 协议转换          │
                    │  ├─ Tool XML 注入/解析 │
                    │  ├─ p5.js 噪声过滤    │
                    │  ├─ SSE 修复          │
                    │  └─ 双层缓存          │

协议转换：将 Anthropic 或 OpenAI 格式的请求转换为 p5js.ai 上游格式（messages + provider + model + deviceId + sessionId）
Tool Use 仿真：将工具定义注入系统提示词为 XML 格式，将上游文本响应中的 <function_calls> XML 块解析回标准 tool_use / tool_calls
噪声过滤：检测并剥离上游自动注入的 p5.js 助手问候语、标题、尾缀推荐
SSE 修复：上游输出的畸形 ddata: / ata: 前缀自动修正为标准 data:
缓存：完整响应缓存，流式和非流式共用，支持 in-flight 合并

注意事项

上游 p5js.ai 会在没有 system 字段时注入 p5.js 助手提示词，客户端显式传 system（Anthropic）或 {"role":"system"} 消息（OpenAI）即可覆盖
流式缓存命中时会重新生成新的响应 ID / 时间戳，并按本地协议重新渲染
默认缓存上限 32 MiB，超过不会截断返回，只是跳过缓存，在 /health 的 cache.oversize_skips 可见
设 RESPONSE_CACHE_MAX_ENTRY_BYTES=0 可取消上限
本服务依赖 p5js.ai 接口可用性，上游变更可能影响使用

致谢

感谢小辣椒的临时邮箱提供注册支持

许可证

MIT License

p5js.ai 2 API — Wrap p5js.ai Free Chat as Anthropic / OpenAI Compatible API

p5js.ai 2 API is a lightweight reverse-proxy adapter that wraps the free chat endpoint at https://p5js.ai/api/ai-chat into both an Anthropic Messages API and an OpenAI Chat Completions API compatible interface.

This means any tool that speaks either protocol — Claude Code, Chatbox, NextChat, LobeChat, one-api, Cherry Studio, etc. — can connect to it as if it were a real Anthropic or OpenAI endpoint.

Key Features

Dual protocol — Serves both /v1/messages (Anthropic) and /v1/chat/completions (OpenAI)
Pseudo tool use — The upstream doesn't support native tool_use; this service injects an XML-based <function_calls> prompt and parses the XML back into proper tool_use / tool_calls format
p5.js noise filter — Automatically detects and strips p5.js assistant greetings, headings, and trailing recommendations injected by the upstream
SSE fix-up — The upstream emits malformed ddata: / ata: prefixes; this service auto-corrects them to standard SSE
Two-tier caching — In-memory + optional Redis second-level cache; streaming and non-streaming share the same completion cache
In-flight deduplication — Concurrent identical requests don't hammer the upstream; followers wait for the leader's result
Connection pooling — Shared httpx async connection pool for high-concurrency scenarios
Proxy support — Supports UPSTREAM_PROXY_URL or standard HTTP_PROXY / HTTPS_PROXY env vars

Project Structure

p5js/
├── main.py              # FastAPI app entry point, route definitions
├── config.py            # Constants, env vars, model list, regex patterns
├── filters.py           # p5.js noise filtering, tool-aware text buffering
├── tools.py             # Tool XML prompt building, parsing, extraction
├── translate.py         # Protocol translation (Anthropic/OpenAI → upstream format)
├── upstream.py          # Upstream HTTP client, SSE parsing, live capture
├── render.py            # Artifact → Anthropic/OpenAI JSON/SSE rendering
├── stream.py            # Live stream handling (with/without tools, with cache integration)
├── response_cache.py    # Two-tier cache system (memory + Redis)
├── tests/               # Test suite
│   ├── test_response_cache.py
│   └── test_filters.py
├── Dockerfile
├── docker-compose.yml
├── requirements.txt
├── start.sh
├── .env.example         # Environment variable template
└── README.md

Quick Start

Local

./start.sh

The first run automatically creates a venv, installs dependencies, and listens on http://127.0.0.1:18185.

Docker

# Build and run
docker build -t p5js2api:latest .
docker run --rm -p 18185:18185 p5js2api:latest

# Or with docker compose (includes Redis)
docker compose up -d --build

The default compose config spins up a local Redis instance. Port is controlled via the P5JS2API_PORT env var (default 18185).

Supported Models

Model	Notes
`claude-opus-4-7`	Latest flagship
`claude-opus-4-6`
`claude-opus-4-1` / `claude-opus-4-1-20250805`
`claude-opus-4-20250514`
`claude-sonnet-4-6`
`claude-sonnet-4-5` / `claude-sonnet-4-5-20250929`	Default
`claude-sonnet-4-20250514`
`claude-haiku-4-5` / `claude-haiku-4-5-20251001`	Lightweight & fast

API Endpoints

Path	Method	Description
`/health`	GET	Health check — returns service status, cache stats, upstream config
`/v1/models`	GET	Model list (Anthropic format)
`/v1/messages`	POST	Anthropic Messages API, supports `stream`
`/v1/chat/completions`	POST	OpenAI Chat Completions API, supports `stream`

The service does not validate API keys — any non-empty string is accepted.

Usage Examples

Claude Code (Anthropic Protocol)

export ANTHROPIC_BASE_URL=http://127.0.0.1:18185
export ANTHROPIC_AUTH_TOKEN=dummy
export ANTHROPIC_MODEL=claude-opus-4-7
export ANTHROPIC_SMALL_FAST_MODEL=claude-haiku-4-5

claude

OpenAI Python SDK

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:18185/v1",
    api_key="sk-dummy",
)

resp = client.chat.completions.create(
    model="claude-sonnet-4-5",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True,
)
for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="", flush=True)

Third-party Tools (Chatbox / NextChat / LobeChat / one-api / Cherry Studio etc.)

Base URL: http://127.0.0.1:18185/v1
API Key: Any non-empty string (e.g. sk-dummy)
Model: Pick one from the "Supported Models" table above

curl

# Anthropic protocol
curl http://127.0.0.1:18185/v1/messages \
  -H 'content-type: application/json' \
  -H 'x-api-key: dummy' \
  -d '{"model":"claude-sonnet-4-5","max_tokens":1024,"messages":[{"role":"user","content":"hi"}]}'

# OpenAI protocol
curl http://127.0.0.1:18185/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"claude-sonnet-4-5","messages":[{"role":"user","content":"hi"}]}'

Caching

The service enables full-response caching by default:

In-process memory cache first (LRU, default 256 entries)
Configure RESPONSE_CACHE_REDIS_URL to upgrade to memory + Redis two-tier caching
Concurrent identical cache misses are in-flight deduplicated — followers wait for the leader
stream=true and stream=false share the same completion cache

Environment Variables

See .env.example for the full list. Key settings:

Variable	Default	Description
`RESPONSE_CACHE_ENABLED`	`true`	Enable/disable caching
`RESPONSE_CACHE_TTL_SECS`	`300`	Cache TTL for plain requests (seconds)
`RESPONSE_CACHE_TOOL_TTL_SECS`	`120`	Cache TTL for tool-use requests (seconds)
`RESPONSE_CACHE_MAX_ENTRY_BYTES`	`33554432`	Max cache entry size in bytes (0 = unlimited)
`RESPONSE_CACHE_REDIS_URL`	—	Redis URL; set to enable second-level cache

Response Headers

X-Proxy-Cache: HIT | MISS | BYPASS | DISABLED
X-Proxy-Cache-Source: memory | redis | inflight | live

Bypass Cache

Either of:

Request header X-Proxy-Cache: bypass
Request header Cache-Control: no-cache

Proxy Configuration

Two proxy options inside the container:

Explicit UPSTREAM_PROXY_URL
Standard env vars HTTP_PROXY / HTTPS_PROXY / NO_PROXY

The /health endpoint's upstream.proxy_configured field shows whether a proxy is detected.

How It Works

┌─────────────┐     ┌──────────────────┐     ┌─────────────┐
│  Client      │────▶│  p5js.ai 2 API   │────▶│  p5js.ai    │
│ (Claude/     │◀────│  (this project)   │◀────│  upstream   │
│  OpenAI)     │     │                   │     │             │
└─────────────┘     └──────────────────┘     └─────────────┘
                    │  ├─ Protocol translation      │
                    │  ├─ Tool XML inject/parse      │
                    │  ├─ p5.js noise filtering      │
                    │  ├─ SSE fix-up                 │
                    │  └─ Two-tier caching            │

Protocol translation: Converts Anthropic or OpenAI format requests into the p5js.ai upstream format (messages + provider + model + deviceId + sessionId)
Tool use emulation: Injects tool definitions as XML into the system prompt, parses <function_calls> XML blocks from upstream text responses back into standard tool_use / tool_calls
Noise filtering: Detects and strips p5.js assistant greetings, headings, and trailing recommendations auto-injected by the upstream
SSE fix-up: Auto-corrects malformed ddata: / ata: prefixes to standard data:
Caching: Full-response caching shared between streaming and non-streaming, with in-flight deduplication

Caveats

The upstream p5js.ai injects a p5.js assistant prompt when no system field is present. Passing system (Anthropic) or a {"role":"system"} message (OpenAI) overrides it.
Cache hits re-generate new response IDs/timestamps and re-render per the local protocol.
Default cache entry limit is 32 MiB. Oversized entries are not truncated — they're simply not cached. Check /health → cache.oversize_skips.
Set RESPONSE_CACHE_MAX_ENTRY_BYTES=0 to remove the limit.
This service depends on p5js.ai availability. Upstream changes may affect functionality.

Acknowledgements

Special thanks to 小辣椒的临时邮箱 (Xiaolajiao Temp Mail) for registration support

License

MIT License

Name		Name	Last commit message	Last commit date
Latest commit History 1 Commit
tests		tests
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
Dockerfile		Dockerfile
README.md		README.md
config.py		config.py
docker-compose.yml		docker-compose.yml
filters.py		filters.py
main.py		main.py
render.py		render.py
requirements.txt		requirements.txt
response_cache.py		response_cache.py
start.sh		start.sh
stream.py		stream.py
tools.py		tools.py
translate.py		translate.py
upstream.py		upstream.py

Folders and files

Latest commit

History

Repository files navigation

p5js.ai 2 API

p5js.ai 2 API — 将 p5js.ai 免费接口包装为 Anthropic / OpenAI 兼容 API

核心特性

项目结构

快速开始

本地运行

Docker 部署

支持的模型

接口一览

使用示例

Claude Code（Anthropic 协议）

OpenAI Python SDK

第三方工具（Chatbox / NextChat / LobeChat / one-api / Cherry Studio 等）

curl

缓存系统

环境变量

响应头

跳过缓存

代理配置

工作原理

注意事项

致谢

许可证

p5js.ai 2 API — Wrap p5js.ai Free Chat as Anthropic / OpenAI Compatible API

Key Features

Project Structure

Quick Start

Local

Docker

Supported Models

API Endpoints

Usage Examples

Claude Code (Anthropic Protocol)

OpenAI Python SDK

Third-party Tools (Chatbox / NextChat / LobeChat / one-api / Cherry Studio etc.)

curl

Caching

Environment Variables

Response Headers

Bypass Cache

Proxy Configuration

How It Works

Caveats

Acknowledgements

License

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages