开源 Claude Opus 5.5 + JEV 生产级 agent 技术栈 20 个仓库合集

qortex @28qortex原推发布于 2026.09.29

在 X 上打开原推

Holy sh*t, someone just open-sourced the complete Claude Opus 5.5 + JEV production stack 20 open-source GitHub repositories with 340K+ combined stars to build production agents with JEV + Claude Opus 5.5, not fragile demo wrappers. Decision layers. Protocols. Sandboxes. AST parsers. Temporal graphs. Memory. Evals. Observability. Security. 01 MCP SDK - official standard protocol connecting structured tools to Claude Opus 5.5 ▸ http://github.com/modelcontextprotocol/typescript-sdk 02 FastMCP - ultra-fast Model Context Protocol server framework for high-throughput tool calling ▸ http://github.com/PrefectHQ/fastmcp 03 Anthropic Quickstarts - reference agent harnesses for computer use, code execution, and multi-modal tasks ▸ http://github.com/anthropics/anthropic-quickstarts 04 LangGraph - cyclic state machine orchestrator where Jev evaluates Choice at conditional routing edges ▸ http://github.com/langchain-ai/langgraph 05 LiteLLM - unified proxy routing 100ms scalar checks to Jev ($0.042/M) and deep reasoning to Opus 5.5 ▸ http://github.com/BerriAI/litellm 06 BAML - typed prompt compiler enforcing rigid schema boundaries on streaming Opus outputs ▸ http://github.com/boundaryml/baml 07 Repomix - packs entire Git repositories into single XML-tagged files for clean Opus context prefill ▸ http://github.com/yamadashy/repomix 08 ast-grep - lightning-fast structural AST search and rewrite engine verifying code trees before commit ▸ http://github.com/ast-grep/ast-grep 09 Docling - IBM layout-aware parser turning complex enterprise PDFs and tables into clean Markdown ▸ http://github.com/docling-project/docling 10 Graphiti - dynamic temporal knowledge graph tracking how repo facts and entity relations evolve over time ▸ http://github.com/getzep/graphiti 11 Letta - stateful agent OS managing working context and paging memory in and out of Opus without context rot ▸ http://github.com/letta-ai/letta 12 E2B - sub-100ms Firecracker microVM cloud sandboxes executing untrusted Python and Bash diffs safely ▸ http://github.com/e2b-dev/E2B 13 Daytona - open-source development workspace orchestrator spinning up sovereign agent devboxes ▸ http://github.com/daytonaio/daytona 14 Temporal - resilient long-running workflow engine ensuring multi-hour Opus loops survive network crashes ▸ http://github.com/temporalio/sdk-typescript 15 Browser-Use - autonomous Chromium browser control where Jev gates destructive DOM actions ▸ http://github.com/browser-use/browser-use 16 Playwright - headless browser test suite verifying web rendering and DOM state before task signoff ▸ http://github.com/microsoft/playwright 17 Semgrep - static AST security analysis blocking hardcoded secrets, SQL injections, and eval exploits ▸ http://github.com/semgrep/semgrep 18 Promptfoo - automated LLM evaluations, regression benchmarks, and security red-teaming in CI/CD ▸ http://github.com/promptfoo/promptfoo 19 DeepEval - production LLM evaluation framework testing hallucination metrics and G-Eval rubrics ▸ http://github.com/confident-ai/deepeval 20 Phoenix - open-source AI observability visualizing agent trace DAGs, latency jitter, and token spend ▸ http://github.com/Arize-ai/phoenix the loop: collect codebase state → Jev evaluates sub-token Choice, Score, or Noul → code enforces policy barrier → Claude Opus 5.5 executes deep architectural synthesis → sandboxed tool runs diff → AST & eval test verifies output → commit state A frontier model without a decision layer is an expensive guessing machine. Wiring Claude Opus 5.5 to Jev and open-source infrastructure turns stochastic AI into an unbreakable software engineer. Save this, then build a second brain for your business ⭣

这条是引用着另一条写的：x.com/28qortex/status/2103949670435492020