Agent frameworks & infrastructure
Agent frameworks & SDKs
12,596 agents · most starred first
Agent frameworks & SDKs, page 50 of 126
Language-agnostic best practices for building reliable agentic loops. Pillar-by-pillar guide to goals, persistent state, verification gates, guardrails, and anti-patterns. Zero dependencies.
Reference implementation of durable execution patterns (signals, queries, saga pattern, idempotency, replay-safe determinism) using Temporal, applied to an agentic triage workflow.
给 DSH agent 的插件库"起飞前检查单"——任务型请求时自动预检插件库并注入 Top-K 适用插件提示,让插件库利用率可预期、不靠运气。Pre-flight plugin-library check for DSH agents — task-type requests trigger a Top-K hint of suitable plugins injected into the runtime context, making plugin usage predictable instead of opportunistic.
You ask. ANT Acts.
Code samples connecting Python agentic frameworks (LangChain, LlamaIndex, CrewAI, ...) to LucidLink via the Python SDK and MCP server
Spec-first, embeddable agent harness runtime - a control plane for loop-engineered AI agents. Governed tool calls through one gateway, one normalized event stream, durable sessions.
Go library that cleans malformed/polluted LLM output (think blocks, tool-call wrappers, fences, preamble/postamble, truncated JSON) — stdlib only, thread-safe
Capability-first vertical agent framework with policy, approvals, workflows, tracing, evals, and an interactive visual lab.
DeepSeek Harness 的 Python 重实现:插件内核 + 事件溯源 + 五段工具管线 + wanter(water+ant)物理经验层——纯梯度 0% 逃逸 vs wanter 100%,路径复用 ≈30× 加速
Agentic AI framework for existing frontend apps — turn natural language into real product actions with typed tools, skills, LangGraph, RAG, memory, and framework-agnostic integrations.
EiHAI builds
Workflow-net runtime for DeepSeek Harness: resource-aware concurrency, native loops and fan-out, static soundness checking before a plan becomes durable, and process mining over its own event log.
A lightweight agent framework for production runtime. Cycle-based execution with pluggable LLM backends, tool dispatch, memory compression, and distributed scheduling.
基于 Claude Code 架构精简的 AI Agent 框架 | 7层架构 Agentic Loop + 6个内置工具 + 三级权限控制 + 动态System Prompt | 可以更加详细的学习工程级架构,也可以自定义内置工具使用
Go engine for DAG-based event routing: fan-in/out, backpressure, retry/DLQ, and observability built in.
A persistent agent runtime and its official CLI. Three packages: @echo-agent/core (runtime/SDK, zero runtime dependencies), echo-agent (CLI and interactive TUI), and echo-coding (the coding-agent product).
losp - a streaming programming language specifically designed for LLM structured metacognition. It is influenced by Lisp, FORTH, brainfuck, Perl but with a healthy dose of novelty.
Natural-language tool calling for LLM agents — Python implementation of the NLT paper (arXiv:2510.14453). Boosts tool-calling accuracy on weak/free-tier models by up to 14%.
Stateful conversational AI agent built with LangGraph, SQLite, and Streamlit. Features dynamic ReAct tool-calling loops, multi-thread session persistence via SqliteSaver, and 4-bit NF4 quantized local inference (Qwen 2.5) optimized for low-VRAM deployment.
面向长任务的 Agent Harness 与受控技能自进化框架。基于 Asyncio ReAct 与 stdio MCP,实现任务状态驱动的上下文治理、业务断言 Oracle、有界修补与防退化版本晋升。
A tiny but mighty agent that runs in your browser, on your phone, or at the edge.
梁神按需模式 / LiangShen Lazy — DeepSeek Harness 的按需工具隔离 preset:默认极简,需要什么能力才临时开放什么,用完自动收回。A minimal-by-default, capability-on-demand agent preset for DeepSeek Harness.
AI-powered framework for building voice-enabled, context-aware assistants with LangChain, OpenAI GPT, FAISS vector search & Vosk speech recognition. Modular design for document retrieval, conversational memory & workflow automation.
A working AI agent built in C from scratch — no frameworks, no SDK. ReAct loop, tool system, filesystem sandboxing, 200+ tests.
Self-hosted Python agent runtime for Linux with Telegram, Discord and CLI; durable conversations, tasks, workspaces, memory, tools and delivery through an AHC-managed Codex backend. Pre-release.
Universal headless AI Assistant engine with extensible tool execution and custom relay support.
治理驱动的企业决策运行时系统
A production-grade, fully asynchronous C++23 agentic AI runtime that orchestrates ReAct (Reason-Act-Tool) loops with local and cloud LLMs. Zero external dependencies — JSON, coroutines, epoll/kqueue reactor, MCP transport, and HTTP client are all hand-written in modern C++23.
Tiny header-only C++20 AI agent framework for edge and embedded: DeepAgent sub-agents, middleware, MCP, SSE streaming, RAG vector stores (Qdrant, Weaviate, Milvus, Redis, Chroma), tracing to Phoenix/Langfuse. 7 MB binaries, 2 MB RSS, runs on Raspberry Pi and Jetson with llama.cpp.
TrustChain Agent OS — Trust-native protocol layer for AI agents. Based on TrustChain (Otte, de Vos, Pouwelse — TU Delft).
TypeScript SDK for building multi-agent LLM applications — event-sourced runtime, plugin-based tools, React client
Universal, credential-free operating framework for AI agents.
Universal Brain x Harness collaboration framework — CLI-first, adapter-first, local-first. 让最强的模型思考,让你顺手的 Agent 干活。
Local-first AI Agent framework. DeepSeek-compatible (deepseek-v4-pro / deepseek-v4-flash), multi-model (OpenAI/Claude/GLM/Xfyun), Core Loop Kernel, sub-agent multimodal delegation, Windows desktop app.......本地优先的 AI Agent 框架,DeepSeek 兼容。
Clawra Engine: 让每个AI都有自我感受 — 自我感知 × 自主进化 × 跨实例连续性
Modular AI agent runtime for explainable, stateful, and bounded workflows with explicit module, capability, and execution boundaries.
Kotlin Multiplatform library wrapping the DeepSeek API with Ktor: streaming chat with interruption and regeneration, a tool-call pipeline, FIM, image input via inline base64 / URL / Files API, and a message-building DSL.
Free & Open Source AI Agent — Self-Hosted, Self-Learning, No Vendor API. Qwen2.5-7B + LoRA. 35 tools. Built on Islamic Epistemology (IHOS).
gguf granite q4_k_m ollama autonomous-agent tool-use
LLM agent framework built from scratch with TDD: tool calling, memory, RAG, MCP, multi-agent delegation and evaluation.
LLM Agent 服务:LangChain + FastAPI + Vue3,混合检索 RAG 带引用溯源,ReAct 双架构实测择优(98.5% vs 89.7%),语义缓存零误命中,158 条评测 + CI 回归门禁,Docker 公网部署
Minimal Python implementation of DeepSeek Harness: plugin loop, hooks, permissions, goals/plans — one command, zero config
Autonomous local LLM agent with planning, tool execution, and self-reflection loop running fully offline
A Laravel agent harness with swappable loops. Pick how your AI agent thinks, switch with one line, compare which works best.
Client SDKs for the Agent Execution Protocol (AEP), the wire protocol behind Helix
The agent loop as a stateless HTTP service
Composable, stateful AI workflows built with LangGraph, implementing sequential, parallel, conditional, iterative, human-in-the-loop, and multi-agent orchestration through graph-based execution. Models use nodes, edges, schema-defined state, routing, checkpointing, tool orchestration, LLM-driven control flow, and graph compilation.
AgentForge is a hands-on platform for building, testing, and comparing modern AI agent frameworks. It features real-world examples, multi-agent workflows, and benchmarking tools to evaluate performance, scalability, and design trade-offs across leading solutions.
Simple, local AI agents for Banking, Retail and Insurance, built with Ollama, CrewAI and LangChain
Build design patterns for autonomous AI agents with orchestration, memory, fallback, tools, and multi-agent coordination
Production-ready LangChain agent template with RAG, tool use, memory, and vector store integration
A digital organism in Rust. Life is cheap, thought is expensive: a pure, deterministic brainstem decides when the LLM may think, and records why it didn't.
A small terminal agent harness: bring your model, use a few tools, keep execution under your control.
Open source AI agent harness & agentic AI runtime in TypeScript. 6-phase state machine, 24 tools, human-in-the-loop permissions, MCP server, and composable agent APIs. Zero dependencies.
A Java Spring Boot agent harness powered by JEV and LLM intelligence.
😄ღ(´・ᴗ・`):A local-first, self-built Agent Harness from scratch. 8-layer enforced architecture (kernel ↔ policy ↔ provider ↔ shell), event-sourced resumable state, pluggable agent policy, BYOM via LiteLLM. Desktop (Tauri v2) + CLI + OpenAI-compatible API.
AYGENT is an Open Source AI Agent Harness with built-in security, memory, cron, and parallel threads.
Self-hosted framework for LLM agents, built around plugins: agents defined in YAML, run in the browser, from the CLI or through an OpenAI-compatible API.
ECHO-CLAW is an AI Remote Viewing training framework built on a Monitor–Viewer architecture, blind-target separation, structured memory, and cumulative learning. It combines theoretical design with practical OpenClaw-based workflow for disciplined, target-blind AI session work and long-term experimental development.
Embeddable AI agent core: OpenAI-compatible providers, tools, skills, MCP, memory, sessions, an ACP-compatible sidecar protocol, and a standalone CLI. Zero runtime deps.
ReAct agent loop (Thought, Action, Observation) in synchronous Rust with no async runtime or I/O, so it builds for wasm32 and runs in browsers, edge workers and plugins.
把智力从模型搬进载体的本地 AI 助手:换模型,助手不变(记忆/人格/工具/安全边界都不丢)。用本地 4B 跑出远超参数量的效果 —— 状态在代码层真的改变行为(工具表 6→0),不是写进提示词劝它。100% 本地运行。
基于 langgraph 图工作流,dsh cordis 插件思想框架的通用agent内核
ACSE — a sovereign cognitive agent built on SQCIG (Schematized Quadrupole Causal Ignition Graph): an A-D-O-G cognitive quadrupole wired as a living causal graph, in a single Zig binary. Every claim measured, never declared.
Async-first Python runtime for building provider-agnostic LLM, agent, tool-calling, streaming, caching, and structured-output workflows across OpenAI, Anthropic, and Google Gemini.
A minimal context kernel for AI agents
自托管的开源多模型 AI 智能体应用:ReAct 循环 + 计划执行 + 长期记忆 + 上下文压缩 + 子代理 + 人工审批 + MCP 扩展 + 流式思考 + Trace 回放 + 内置评估框架(25 任务 100% 通过)。14 个内置工具,五层安全防护,Flask + SQLite 轻量架构,双击即可部署,数据完全私有。
Reliable, embeddable TypeScript agent runtime with durable runs, tool safety, recovery, evidence, and bounded multi-agent coordination.
Open-source agent harness for knowledge work: long-running AI agents with MCP tools, managed context, subagents and self-improvement in one lean TypeScript-on-Bun binary.
AgentOS — local autonomous agent runtime: plan, execute, verify, review and self-correct engineering tasks (TypeScript core + CLI + Next.js dashboard)
Self-hosted AI agent workbench - long-term memory, knowledge base, full toolchain. Desktop + mobile web + QQ/WeChat, bring your own model. With you, I can do anything. Run by me, Nis.
AgentForge: Open-source AI Agent Platform — Build, Deploy, and Orchestrate Intelligent Agents at Scale with Next.js 15, TypeScript & Docker.
AI Agent:手搓 · 框架 · 企业级。Tool calling, memory, planning, MCP, evaluation, guardrails, concurrency & resume.
👾 下一代透明智能体架构 | Next-Gen Transparent Agent Architecture 🔍 全行为审计 | 🛡️ 两段式安全调用 | 🧠 双水位记忆 | ⏰ 心跳任务 📊 P0 级事故率降低 80% | 兼容 OpenClaw + Claude Code 技能生态
A modern, plugin-driven TypeScript AI Agent framework. 44 core files, strict TS, zero boilerplate.
LLM SDK for Go: one typed API across Anthropic, OpenAI and Gemini, and the agent loop around it.
Build and operate agentic AI systems, workflows, RAG pipelines, models, tools, and runtime infrastructure from a unified control plane. Deploy across cloud, VPC, or self-hosted environments with built-in versioning, observability, governance, and release management, so teams can move from prototype to production without rebuilding the stack.
KAIROS is a lightweight, self-improving AI agent
Frontier Agentic Language
Lightweight, provider-agnostic agentic tool-use loop for .NET. Define your tools; the engine runs the model/tool loop. A from-scratch reference implementation.
Hades — runtime agentic local-first com multi-provider, Claude/Codex/OmniRoute/xAI adapters, browser/computer use, Company OS/Growth sandbox, builders, MCP/plugins/skills lifecycle, approvals e hardening enterprise
Specton — 고객 요구사항 기반 시스템 기능·데이터 모델 자동 설계 Agent (팀 설계로)
SWIFT: Synthesizing Workflows via Few-shot Transfer. Amortized LLM agent workflow synthesis from reusable structural priors.
A small, readable Python toolkit for building AI agents that safely work with your files. Runs free on local models via Ollama, or with OpenAI and Anthropic.
Spring-style TypeScript framework for mission-critical LangGraph AI agents — @Application DI, Path Scope Flow DSL, transactional LIFO sagas, FinOps budgets, MCP/RAG/A2A & unit testing.
Self-hosted AI server and agent framework. Pair your phone, browser or desktop with your own computer peer to peer over WebRTC, and run models, agents and web apps on hardware you control. Source-available. Your own AI Automation software, with native build instructions for macOS, Windows, WSL (Ubuntu), and Docker. iOS, Android, Store Apps separate
Minimal AI agent template in Python (Claude) — agentic tool-use loop, no framework
A learning-example Spring Boot application demonstrating core LangChain4j concepts against Anthropic's Claude models: plain AI Services, chat memory, tool/function calling, and Retrieval-Augmented Generation (RAG).
A Guile 3 implementation of an AI agent following the ampcode.com pattern for tool-based interactions
Maps external tool surfaces (e.g., MCP servers, APIs, CLIs, SDKs) into a persistent capability registry agents can load and act on.
Fixes DSH's silent failure where a tool call never runs and its markup (bare invoke/parameter tags, antml:-prefixed, or a stray 'court') leaks into the reply as text: blocks damaged calls, self-heals in-turn, and after a fork voids the damaged output on the model-visible surface. Zero deps.
The authoritative standard and developer skill for DeepSeek Harness (DSH 0.2.0-rc.2) & Cordis plugin development.
Enterprise-grade AI agent engine. Lightweight, secure, MCP-native. Built in Rust.
Agent-native, governed AI runtime for Rust, Python, and TypeScript — native providers, MCP, tools, routing, memory, budgets, audit, and containment.
Ouvrier is a Go framework for turning an API, webhook, cron job, or stream into reliable, governed agentic middleware by simply declaring pipelines such as From -> Pipe -> Outcome.
gradio region:us
streamlit region:us
streamlit region:us