LiteLLM (YC W23)

Thanks to my personal fav @LiteLLM, switching to a faster api model was eassssy. So been using @AnthropicAI Haiku

great guide on using @openclaw (ClawdBot) + @LiteLLM + @nebiustf  👍

@NirantK Looking into it

.@LiteLLM has a bug on their Gemini LLM caching which is leading to 2-10x costs, we raised a PR to fix 3 days ago. How do we get someone to take a review? PR link below from our team

Boy, have we got news for you! 🧑‍🍳 The @openwork_ai team is happy to announce: - Amazon Bedrock integration, contributed by our friends at @awscloud (thanks guys!) - Native @deepseek_ai integration - Integration with @openrouter and @LiteLLM! Here's how what it looks like >>

first, pinky, we will use @DSPyOSS and @LiteLLM proxy server with custom guardrails to translate any api call into spanish... ...then, we try to take over the world!

@LiteLLM v1.77.4-nightly brings support for the new Anthropic web fetch tool The web fetch tool allows LLMs to retrieve full content from specified web pages and PDF documents Get started here: docs.litellm.ai/docs/completio… List of improvements here:

@aimlapi 🤝

If you want to quickly and efficiently integrate an LLM into your "Internet Of Agents" hackathon project, @LiteLLM is your best friend. → Switch between any LLM provider with a single line of code → Experiment and deploy models without rewriting your functions → Keep code clean, modular, and fully reusable

In startups, you are constantly migrating. From fast & simple to complex, custom and scalable. In the last 4 months we’ve migrated: 1. NextJS on Vercel -> React-Router on DigitalOcean 2. Custom agent code & @LiteLLM -> @pydantic AI 3. @digitalocean -> @awscloud 4. @langchain -> custom RAG

I once again learned something new from @kweinmeister. @LiteLLM looks like a nice proxy in front of multiple LLMs in @googlecloud Vertex AI. Karl shows how you could invoke this directly, in Cline or Roo Code, and as a persistent background service.

@LiteLLM v1.77.0-nightly brings major security upgrades - from this release there will be 0 CVSS Vulnerabilities with a 4.0 or higher score We plan on guaranteeing this baseline for all future releases of LiteLLM Images Other improvements on this release 👇

@nvidia Integration made simple ⚡ Thanks to @LiteLLM , we support: - 100+ LLM providers - Your own custom chatbots - Seamless testing across different AI systems

Want to use models from Google's Vertex AI Model Garden with an OpenAI-compatible API? My new video shows how to set up @LiteLLM as a local proxy to do just that. Simplify your workflow and call models like Qwen, DeepSeek, and more through a unified interface. #AI #LLM #VertexAI #LiteLLM #GoogleCloud

LangChain is a 10,000 line wrapper for a 3-line API call. Many layers of abstraction. x dependencies. To do what? Call LLMs. Deleted everything. Switched to @litellm. Same functionality. Actually faster. And I can see what's happening. The "Agent Framework" grift is selling you complexity as intelligence. They're not making AI easier. They're making consultants richer. Your "production-ready enterprise solution" is someone's weekend project wrapped in abstract base classes. Stop LARPing as a software architect. You're calling an API.

@LiteLLM v1.76.1-nightly brings support for the new OpenAI gpt-realtime model family You can now use remote MCP servers, image inputs with gpt-realtime Other improvements on this release: Fix Next.js Security Vulnerabilities in UI Dashboard

You can use @xai Grok-4 for Claude Code simply with @LiteLLM 👋🏻 [Terminal 1] 1. export keys 2. run claude --model grok-4 [Terminal 2] 1. edit config.yaml ``` general_settings: master_key: your-master-key proxy_user_auth: false model_list: - model_name: grok-4 litellm_params: model: xai/grok-4 api_key: "os.environ/XAI_API_KEY" ``` 2. export keys 3. run litellm --config config.yaml

@LiteLLM v1.76.1-nightly brings support + cost tracking for Google gemini-2.5-flash-image models From this release you can use gemini-2.5-flash-image-preview on Google AI Studio and Vertex AI Other improvements on this release 👇

Claude Code 为什么这么好用？背后的“魔法”是什么？是它精心设计的提示工程和上下文管理，一起看看 👇 Claude Code 为什么“就是好用”？ 许多开发者对它“几乎无脑好用”的体验赞不绝口。作者通过设置 @LiteLLM 代理，拦截 Claude Code 与 Anthropic API 之间的通信，分析了它的工作原理。核心发现是：Claude Code 的成功不在于模型本身有多聪明，而是通过大量的 标签和系统化的上下文管理，让模型始终保持专注，避免“跑偏”。 核心机制：无处不在的“提醒标签” Claude Code 的“秘诀”在于它在整个工作流程中频繁使用 标签。这些标签就像是给模型的“便签”，不断提醒它当前的任务和约束，出现在： · 系统提示（system prompts） · 用户消息（user messages） · 工具调用结果（tool results） · 甚至是 bash 命令输出中 例如，当你更新一个待办事项（todo），系统可能会插入4个以上的提醒标签，确保模型不会忘记任务目标。这种“偏执”设计让 Claude Code 在执行复杂任务时也能保持高度专注。 工作流程：步步为营 Claude Code 在开始工作前会做一系列准备工作，确保任务执行的精准性： · 上下文预加载：在正式工作前，Claude会先总结对话（50字以内）、提取主题（2-3词）、检查命令是否安全，防止注入攻击。 · 子智能体分派：对于复杂任务，Claude会启动专门的“子智能体”（sub-agents），每个子智能体专注于特定任务，减少主模型的负担。 · 持续提醒：通过 标签，系统在每个步骤都注入上下文，确保模型不偏离轨道。 这种设计就像给模型装了个“导航系统”，每走一步都有明确的指引。 安全第一：命令注入检测 Claude Code 在执行 bash 命令时非常谨慎。它会分析命令的前缀，判断是否存在注入风险（比如恶意代码）。如果检测到可疑命令（如 git statusls``），系统会标记为 “command_injection_detected”，并要求用户手动确认。这种机制大大提升了安全性。 子智能体的巧妙设计 Claude Code 会根据任务复杂度动态调整上下文。比如，子代理在执行简单任务时不会使用待办列表工具（TodoWrite），以保持高效。但如果任务变得复杂，系统会通过 标签提醒子代理使用待办列表，确保任务分解和跟踪。这种“按需注入”的策略让子代理既高效又灵活。 为什么只有 Anthropic 这么做？ 博客提出了一个有趣的问题：为什么 标签在 Claude Code 中用得如此频繁，而其他 AI 系统很少见？作者认为这可能与 Anthropic 的模型训练方式有关，但具体原因尚不清楚。不过，这种“强迫性专注管理”显然是 Claude Code 成功的关键。 对开发者的启发 几个可借鉴的模式，供开发者构建自己的 AI 智能体时参考： · 提前加载上下文：在任务开始前总结对话、提取主题，确保模型理解任务背景。 · 频繁使用提醒：通过类似 的标签，在整个流程中不断强化目标。 · 嵌入安全机制：在命令执行前加入验证和权限检查，降低风险。 · 分层智能体设计：用主循环控制多个专注子智能体，根据任务复杂度动态调整上下文。 这些方法的核心是：通过工程化的提示和上下文管理，让 AI 保持专注，而不是单纯依赖模型的“智能”。

@LiteLLM v1.75.10-nightly brings support for using VertexAI @Alibaba_Qwen models From this release you'll be able to use this model family + LiteLLM will track the cost per request Get started here: #vertexai-qwen-api" target="_blank" rel="nofollow noopener">docs.litellm.ai/docs/providers… Other improvements on this release:

@LiteLLM v1.75.9-nightly brings major improvements to our @datadoghq LLM Observability integration. These improvements are great for AI Platform Teams deploying LiteLLM x Datadog into production. Datadog-related improvements in this release 👇