Предложить идею

Building agents: from model API to evals

For engineers new to agent work, in build order: model API, first agent, tools over MCP (the open protocol for agent tools), data, context and memory, a framework, evals, tracing, sandboxing. Skills run in Claude Code or Codex, plugins in Claude Code. Your agents need a model API key, usually paid.

13 ресурсовОбновлено 2026-10-12

  1. LLMCLITry models from the terminal: OpenAI built in; Claude, Gemini and local models via plugins. Python; API keys for hosted.simonw/llm12.6k
  2. Claude APISkillModel API: tool use, streaming, caching and the agent loop in 7 SDK languages plus cURL. Needs an Anthropic API key.anthropics/skills182k
  3. Agent SDK DevPluginFirst agent: /new-sdk-app scaffolds and checks a Claude Agent SDK app in Python or TypeScript. Claude Code; API key.anthropics/claude-code151k
  4. MCP BuilderSkillExpose tools: design and build an MCP server in TypeScript or Python (FastMCP), then write eval questions to test it.anthropics/skills182k
  5. MCP InspectorCLIDebug MCP: list and call a server's tools in a web UI, terminal UI or JSON CLI for CI smoke tests. Needs Node.js.modelcontextprotocol/inspector11.1k
  6. MCP Toolbox for DatabasesMCPConnect data: prebuilt SQL tools or only queries you define in tools.yaml; Python, JS, Go, Java SDKs. Needs a database.googleapis/mcp-toolbox16.7k
  7. Chroma MCP ServerMCPRetrieval: store documents and search by vector, full text or metadata. Runs locally with uvx; Chroma Cloud optional.chroma-core/chroma-mcp599
  8. Agent Skills for Context EngineeringSkillContext design: compression, tool design, memory and multi-agent patterns, plus LLM-as-judge. Not tied to one platform.muratcankoylan/Agent-Skills-for-Context-Engineering18.1k
  9. Graphiti MCP ServerMCPMemory: keep facts and episodes in a temporal knowledge graph. Experimental; needs Docker and an LLM API key.getzep/graphiti31.8k
  10. LangChain SkillsSkillOther framework: LangChain, LangGraph and Deep Agents quickstarts, persistence, human approval. Early stage; model key.langchain-ai/langchain-skills1.3k
  11. promptfooCLIEvals and red-teaming: compare prompts and models in bulk, scan for vulnerabilities, run in CI. Needs provider keys.promptfoo/promptfoo26.1k
  12. MLflow SkillsPluginTracing: add MLflow Tracing to Python or TypeScript apps, debug failed traces, build scorers. MLflow 3.8+; Claude Code.mlflow/skills81
  13. Anthropic Sandbox Runtime (srt)CLISandbox: limit an agent's or MCP server's file and network access, no containers. macOS, Linux; Windows alpha. Preview.anthropics/sandbox-runtime5.6k
Версия в Markdown (для агентов)