https://happy520ai.github.io/unified-ai-system/ Unified AI System - self-hosted AI gateway and agent control plane Every claim in these pages is reproducible: measured tool rosters, credential-free runs, and verbatim quotations from the projects compared. 2026-09-30T00:44:55.842Z Apache-2.0, Unified AI System contributors https://happy520ai.github.io/unified-ai-system/mcp-security-boundaries.zh-CN.html MCP 安全边界：我们实际检查了什么，哪些仍未被证明 2026-09-30T00:44:42Z 对真实网关进程运行的 23 次攻击探测、用两种不同方法完成的已发布镜像内容审查，以及这两种方法各自止步的界线。每个数字都附上可以重跑的命令。 https://happy520ai.github.io/unified-ai-system/index.zh-CN.html Unified AI System｜面向 Codex、Cursor、Cline 的开源 MCP AI 网关 2026-09-30T00:24:43Z 加固后的 Public Preview：面向 Codex、Cursor 和 Cline 的开源、自托管 MCP 与 AI 网关，提供无需 Provider 的提示增强与 12 个 MCP 工具。 https://happy520ai.github.io/unified-ai-system/ Unified AI System 2026-09-30T00:24:43Z Hardened Public Preview: an open-source, self-hosted MCP and AI gateway for Codex, Cursor, and Cline with provider-free prompt enhancement and reproducible verification. https://happy520ai.github.io/unified-ai-system/mcp-security-boundaries.html MCP security boundaries: what we have actually checked, and what is still unproven 2026-09-29T23:57:09Z Twenty-three attack probes against a real gateway process, two published-image content reviews done by different methods, and the line where neither of them reaches. Bounded evidence with the command to re-run each claim. https://happy520ai.github.io/unified-ai-system/credential-free-evidence.html Credential-free AI gateway: how one run proves no provider was called 2026-09-29T23:45:21Z A usage-report envelope with mode=fake, providerCalled=false and credentialRequired=false, plus the enhanced prompt and preserved original request, from a 14-second local run. https://happy520ai.github.io/unified-ai-system/mcp-startup-timeouts.zh-CN.html MCP 客户端为什么在服务端应答之前就超时 2026-09-29T23:22:31Z 实测：stdio MCP 服务端从进程启动到回答 initialize 落在 6,391 到 8,914 毫秒之间（同一台机器 23 次冷启动），而完整的 15 个工具清单只在其后 3 到 7 毫秒就给出。连接预算几乎全部花在进程启动与模块加载，不是 MCP 工作本身。一条命令即可复现，无需 API Key。 https://happy520ai.github.io/unified-ai-system/codex-mcp-docker-quickstart.zh-CN.html Codex 接入 MCP Server：Docker 60 秒中文教程 2026-09-29T21:43:10Z 使用一条 Docker 命令把 Codex 接入自托管 MCP Server，检查 15 个工具，本地增强提示词并完成无需 API Key 的安全测试。 https://happy520ai.github.io/unified-ai-system/codex-mcp-docker-quickstart.html Run an MCP server in Codex with Docker 2026-09-29T21:43:10Z Connect Codex to a self-hosted MCP server with one Docker command, verify fifteen tools, enhance a prompt locally, and make a credential-free test call. https://happy520ai.github.io/unified-ai-system/mcp-ecosystem-measurements.zh-CN.html 对公开 MCP 生态的测量（中文索引） 2026-09-29T20:12:52Z 向官方注册表登记的 40 个服务端匿名提问：9 个问题、10 次测量。每个数字都从机器可读数据集生成，页面自己声明分母与所请求的协议修订。 https://happy520ai.github.io/unified-ai-system/listing-census.html Where this project is listed, with the URL that proves it 2026-09-29T19:12:56Z <!-- Generated by tools/render-listing-census.mjs from two live probes. Do not edit: re-run node tools/render-listing-census.mjs, and the nightly job runs... https://happy520ai.github.io/unified-ai-system/self-hosted-ai-gateways-compared.zh-CN.html 自托管 AI 网关对照：LiteLLM、Portkey、Agent Router 与 Unified AI System 2026-09-29T18:56:45Z 四个可自托管的 AI 网关，只用各项目自己 README 里的原文来描述：附链接、标注阅读日期，不替别人说它缺什么。 https://happy520ai.github.io/unified-ai-system/self-hosted-ai-gateways-compared.html Self-hosted AI gateways compared: LiteLLM, Portkey, Agent Router and Unified AI System 2026-09-29T18:56:45Z Four self-hostable AI gateways described in their own words: verbatim README lines, licenses and links, read on 2026-09-26, plus the three questions to ask any of them before handing over agent traffic. https://happy520ai.github.io/unified-ai-system/mcp-ecosystem-measurements.html Nine measurements of the public MCP ecosystem 2026-09-29T14:22:02Z 9 questions put to 40 servers in the official MCP registry. 2 agreed to a protocol revision that does not exist. Scripts and three published datasets alongside. https://happy520ai.github.io/unified-ai-system/mcp-session-enforcement.html If an MCP server hands you a session id, does it require it back? 2026-09-29T12:40:21Z The 2026-07-28 revision moves MCP from a stateful initialize handshake toward per-request state, and a lot of gateway work right now is deciding what to do... https://happy520ai.github.io/unified-ai-system/multi-arch-node-modules.zh-CN.html linux/arm64 上 better-sqlite3 抛出 "invalid ELF header"：错误的 CPU 架构 node_modules 陷阱 2026-09-29T12:38:49Z 以多平台方式发布的 Node 镜像，它的 linux/arm64 标签里可能装着 x86-64 原生模块，于是第一次 require() 该包就会失败并抛出 invalid ELF header。2026-09-29 从我们自己镜像的四个已发布版本实测； 不需要 Docker 引擎，也没有执行任何代码。 https://happy520ai.github.io/unified-ai-system/multi-arch-node-modules.html "invalid ELF header" from better-sqlite3 on linux/arm64: the wrong-architecture node_modules trap 2026-09-29T12:38:49Z A Node image published as multi-platform can carry x86-64 native modules inside its linux/arm64 tag, and the first require() of the package fails with... https://happy520ai.github.io/unified-ai-system/mcp-tool-definition-quality.html We ran Glama's open tool-definition rubric against our own MCP server. Here is the part that needs no judge. 2026-09-29T12:38:49Z Glama's Tool Definition Quality Score is an open specification: four stages, of which three are deterministic code and one - stage 3 - is an LLM rubric... https://happy520ai.github.io/unified-ai-system/mcp-package-resolve.html Do the MCP registry's non-npm listings resolve? Measured across pypi, OCI, mcpb, cargo and NuGet 2026-09-29T12:38:49Z 15 of 785 listings across 791 entries could not be resolved at the version the registry declares. The per-family numbers differ a lot, and two of these... https://happy520ai.github.io/unified-ai-system/mcp-npm-installability.html Can you actually install what the MCP registry lists? 196 of 200 sampled npm packages could 2026-09-29T12:38:49Z Short answer: the npm listings in the official MCP registry are in good shape. 98% of the sampled records point at a package that exists on npm and... https://happy520ai.github.io/unified-ai-system/mcp-registry-census.html How many MCP servers are there? 37,013, and what can a client do with them 2026-09-29T12:38:49Z Short answer: there are 37,013 servers in the official MCP registry's default view as of 2026-09-28, and 439 of them (1.20%) declare nothing a client can... https://happy520ai.github.io/unified-ai-system/mcp-registry-installability.html Is a registry-listed MCP server actually installable from the registry? 2026-09-29T12:38:49Z Of the 6 that do carry a package: transports stdio 6; registry types npm 6. The sample is the registry's own default order over the first pages, which... https://happy520ai.github.io/unified-ai-system/mcp-list-cache-hints.html Do MCP servers say how long their tool list may be cached? 2026-09-29T12:38:49Z A gateway has to decide how long to keep a cached tool list. The spec gives servers a way to answer that: ttlMs and cacheScope on a list result. So the... https://happy520ai.github.io/unified-ai-system/mcp-route-headers.html Does an MCP server route by the `Mcp-Method` header? 2026-09-29T12:38:49Z The spec puts Mcp-Method and Mcp-Name on requests so a server can route a body-less GET stream. A POST that already carries a JSON-RPC body does not need... https://happy520ai.github.io/unified-ai-system/mcp-protocol-version-header.html Does an MCP server check the protocol-version header? 2026-09-29T12:38:49Z The 2025-06-18 revision of MCP says a client MUST put MCP-Protocol-Version on every HTTP request after initialize, and that a server SHOULD reject a request... https://happy520ai.github.io/unified-ai-system/mcp-protocol-revision-tolerance.html Will an MCP server agree to a protocol version that does not exist? 2026-09-29T12:38:49Z The Model Context Protocol says a server must answer initialize with a revision it supports, and the revision is not decoration: it decides whether... https://happy520ai.github.io/unified-ai-system/mcp-tools-list-pagination-survey.html Does anyone's `tools/list` actually paginate? A measurement. 2026-09-29T12:38:49Z This exists because issue #177 was closed with an honest caveat: the fix walks a paginated tools/list to the end and refuses to present a truncated... https://happy520ai.github.io/unified-ai-system/mcp-startup-timeouts.html Why an MCP client times out before the server answers 2026-09-29T10:10:07Z A stdio MCP server can spend its whole connect budget before it answers the first handshake. We measured where that time goes - three batches in one day, no API key, one command - and found tool enumeration costs 3 to 7 milliseconds while process boot costs seven seconds. https://happy520ai.github.io/unified-ai-system/security-drill-evidence.html 23-attack security regression, run and logged: what a self-hosted AI gateway defends 2026-09-29T08:03:44Z All 23 attacks defended by the repository's own regression against a live local gateway: cross-tenant cache and audit isolation, header forgery, role checks, budget and rate enforcement, instant revocation, secret redaction, and metrics authorization. No credentials, one run. https://happy520ai.github.io/unified-ai-system/verify-mcp-docker-image.html What a published MCP Docker image actually exposes: 9, 12, then 15 tools 2026-09-29T08:03:44Z We read the shipped source out of eight published Docker tags of our MCP server, verified every layer blob against its digest, and printed the tool roster each image really contains. No Docker daemon, no API key. https://happy520ai.github.io/unified-ai-system/prompt-enhancement.zh-CN.html 自然语言提示词增强 2026-09-29T08:03:44Z 使用确定性、无需 provider 的 AI 网关，把普通自然语言需求整理成更容易检查和执行的提示词。 https://happy520ai.github.io/unified-ai-system/openapi-to-mcp.zh-CN.html OpenAPI 转 MCP：把 REST API 变成受治理的 MCP 工具 2026-09-29T08:03:44Z 把 OpenAPI 3.0 或 3.1 文档中语义明确的每个 operation 生成为 MCP 工具，并置于显式允许清单之后（无法解析的构造会被拒绝，而不是猜测）：缺省即全部拒绝、只读声明、出站请求校验，以及审计边界。 https://happy520ai.github.io/unified-ai-system/openapi-to-mcp.html OpenAPI to MCP: turn a REST API into governed MCP tools 2026-09-29T08:03:44Z Convert OpenAPI 3.0/3.1 operations into MCP tools behind an explicit allowlist; a construct whose semantics cannot be resolved is refused rather than guessed: deny-by-default policy, read-only attestation, SSRF-checked outbound calls, and an audit boundary. https://happy520ai.github.io/unified-ai-system/terminal-first-ai-gateway.html Why an AI gateway should be terminal-first 2026-09-29T08:03:44Z A practical design case for terminal-first AI gateways: stable automation, explicit authority, observable execution, and an optional enterprise console. https://happy520ai.github.io/unified-ai-system/prompt-enhancement.html Natural-Language Prompt Enhancement 2026-09-29T07:50:01Z Turn rough natural-language requests into inspectable prompts with a deterministic, provider-free AI gateway and MCP server.