← Safety

核验标记表示资料完整度，不是产品评分。下列未确认项可能影响选型；请按对应来源逐项确认。The confidence mark describes evidence completeness, not a product score. Unconfirmed items below may affect your choice; check them against the relevant sources.

模型与开放条件Model & openness

这是它的主轴之一：官方 providers 页在册「100+ LLM，一种 OpenAI 格式」，密钥自带（BYO key）； 自托管端点也在这条线上（页内在册 Ollama、vLLM、Llamafile）。release notes 在册 provider 持续新增 （2026-11 条目含 Microsoft 365 Copilot、Azure PTU 容量等）。护栏判定不内置，由挂进来的外部 rail 完成。

This is one of its two main axes: the providers page records '100+ LLMs' behind one OpenAI-format API with bring-your-own keys, and self-hosted endpoints on the same page (Ollama, vLLM, Llamafile). Release notes record providers still being added (the 2026-11 entries include Microsoft 365 Copilot and Azure PTU capacity). No guardrail judgement is built in; it comes from external rails you attach.

运行位置Where it runs

自托管为主：simple-proxy 页在册 pip 安装后 litellm --config config.yaml 起服务，默认 4000 端口； deploy 页在册镜像 ghcr.io/berriai/litellm、Docker 运行示例、Helm / Terraform 与 Postgres 依赖、 健康检查 /health/readiness。首页在册另有托管云用法。生产需数据库这点在 simple-proxy 页明写。

Self-hosting first: the simple-proxy page records pip install then litellm --config config.yaml, default port 4000. The deploy page records the ghcr.io/berriai/litellm image, a docker run example, Helm / Terraform, a Postgres dependency and the /health/readiness check. The site also records a hosted cloud usage path. Production requiring a database is stated plainly on the simple-proxy page.

本地文件Local files

读的是你本机的 config.yaml（模型清单、密钥引用、guardrail 配置），密钥与路由定义都在你自己的环境； 花费与日志写进你自管的 Postgres。它不读你的业务代码工作区，文件面就是配置与数据落库两处。

It reads your own config.yaml (model list, key references, guardrail settings); keys and routing definitions stay in your environment, and spend and logs go into the Postgres you manage. It does not read your application source tree - the file surface is configuration plus where you persist data.

关机后的任务Tasks after shutdown

它是常驻网关进程，请求排队、重试与路由由服务侧处理；跨关机续跑取决于你怎么部署（Helm / VM）。 批处理与异步作业面（如 videos / batches）只在 release notes 里出现过，官方专页本档未逐条核验，记为未知。

It is a long-running gateway process: queueing, retries and routing happen service-side, and surviving a shutdown depends on how you deploy it (Helm or a VM). The batch and asynchronous job surface (videos / batches) has only appeared in release notes; dedicated pages were not verified item by item and are recorded as unknown.

工具与扩展Tools & extensions

官方 MCP 页在册它是「MCP Gateway」，提供「governed access to tools」：server 走 config.yaml 或 UI 配置， 传输支持 HTTP / SSE / stdio，鉴权含 OAuth、API key 与 AWS SigV4，访问按 key / team / organization 控制。 另外有 A2A、Auto Router、logging 与 OpenTelemetry 页在文档导航在册。

The MCP page records it acting as an 'MCP Gateway' offering 'governed access to tools': servers configured through config.yaml or the UI, transports HTTP / SSE / stdio, authentication via OAuth, API keys and AWS SigV4, with access scoped by key, team and organization. A2A, Auto Router, logging and OpenTelemetry pages sit in the documentation navigation.

上下文与记忆Context & memory

上下文由调用方透传，网关自身不做记忆或压缩。虚拟密钥、预算与团队是访问层的边界，不是上下文层； 上下文窗口实际大小由你路由到的模型决定。

Context passes through from the caller; the gateway itself keeps no memory and does no compression. Virtual keys, budgets and teams are access-layer boundaries, not context-layer ones, and the effective window is whatever the routed model gives you.

权限与限制Permissions & limits

访问控制是它的核心：virtual-keys 页在册密钥经 /key/generate 签发，可限模型、预算、速率与团队， 主密钥用于管理，定时轮换属企业版。guardrails 页在册多数护栏在开源侧可用， 动态参数与逐 key 护栏控制需企业版。数据流向取决于你接的 provider；SSO 与审计日志在企业版口径下（simple-proxy 页在册）。 合规认证与第三方审计未核验到官方在册页，记为未知。

Access control is its core: the virtual-keys page records keys minted through /key/generate, model, budget, rate-limit and team scoping, a master key for administration, and scheduled rotation as an Enterprise capability. The guardrails page records that most integrations run in the open-source build while dynamic parameters and per-key rail control need Enterprise. Data flow follows whichever providers you attach. SSO and audit logs sit in the Enterprise framing (simple-proxy page). Compliance certifications and third-party audits were not found on official pages and are recorded as unknown.

适合什么任务Best for

适合要把多家 provider 收成一个 OpenAI 兼容入口、同时按 key / team 做预算与访问治理、并坚持自托管的团队。 不适合只要一个薄 SDK 的场景，也不适合把「厂商背书的托管护栏 + 合规承诺」当必需项的团队。

Fits: teams collapsing many providers into one OpenAI-compatible entry point, governing spend and access per key or team, and insisting on self-hosting. Not a fit: anyone wanting only a thin client SDK, nor anyone for whom a vendor-backed hosted guardrail with compliance attestations is a requirement - rails here are external services you plug in, not a vendor guarantee.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.