Python agents

Nutrition-aware McDonald's China MCP skill: pick the best combo under calorie, protein and coupon constraints.

MCP-сервер для hh.ru: поиск, резюме, отклики, переписка с работодателями и скрытая аналитика по одному аккаунту соискателя через собственный JSON сайта (без официального API). Исходники открыты, некоммерческая лицензия.

Open-source runtime for LLM agents with persistent memory, relationship history, modeled mood, and evolving identity. Carries state across conversations into prompts for Ollama, OpenAI-compatible backends, or Codex. Inspectable state, with opt-in autonomous research mode.

Anime & manga recommendation agent over a Neo4j knowledge graph built from MyAnimeList. Local LLMs (gpt-oss-20b via MLX) for text-to-Cypher with schema and value checks, vector search over synopses, a tool-using agent, a web UI and automatic evals. Code and prompts in Portuguese (pt-BR).

CineAgent：基于 FunRec 增量开发的对话式电影推荐系统，含推荐路线对照、偏好记忆、工具调用及故障降级评测。

SLCA-GRPO: Resolving Cross-Segment Credit Misattribution in Tool-Calling RL — segment-locked credit assignment for tool-calling RL, built on verl.

Self-diagnosis and troubleshooting for Hermes Agent.

LLM agent for local code repository analysis: LangChain tool calling + FAISS RAG + structured output + evals (OpenAI-compatible, DeepSeek by default)

Keyboard-first Kubernetes TUI with an AI agent that works on the same screen, plus an MCP bridge for your editor's assistant. Every cluster write waits for your keystroke.

通用 Agent 内核：在线大脑规划 + 本地执行器，MCP / LangGraph / OpenAI Agents SDK 驱动，支持知识级自升级（技能与全局记忆自动提炼）

受约束的 A 股模拟交易 Agent Runtime：LLM 只读提案 → 确定性风控门 → 人工审批 → 幂等执行 → 事件溯源审计

A short introduction to attention in deep neural networks is provided. The important components of the classical transformer architecture are discussed. Following this, a vision transformer (ViT) is implemented and briefly tested.

Production-ready template for a self-hosted Agent Protocol server: FastAPI + LangGraph + PostgreSQL (+ optional Redis), clean-architecture layering, dual graph-authoring paradigms, staged writes, generative UI, and a CopilotKit demo frontend. E2E-tested with real SDK clients.

A Copilot-style AI chat panel that lives inside Blender. The agent loop runs in Blender's own Python process, executes the model's code against your live scene, and wraps every turn in one Ctrl+Z.

Let AI glasses hand work off to a Hermes agent on your own machine: async task queue + MCP server + adapter, so long-running agent work never times out. 眼镜下单，Hermes 干活。

Free & Open Source AI Agent — Self-Hosted, Self-Learning, No Vendor API. Qwen2.5-7B + LoRA. 35 tools. Built on Islamic Epistemology (IHOS).

A small, plantable (parasitic but only in a good sense kefir) culture for development through self-observation.

autonomous web developer orchestration with python. now has hierarchical RAG using vector databases

Multi-model agent harness with orchestration, Claude Code compatibility, local LLM support, and live agent monitoring.

Official implementation of Every Token Counts: Online Memory Worth for Resource-Constrained Lifelong Agents (NeurIPS 2026 TTCL Workshop).

Zero-trust, risk-aware gateway proxy for the Model Context Protocol (MCP) — enforces identity-scoped tool access, detects schema drift ("rug pull") attacks, scores tool-call risk, and produces a tamper-evident audit trail.

EdgeLab Trading Skills｜给 AI Agent 用的投研与交易复盘工具集 · 11 Skills · CC BY-NC 4.0

Budget-tier LLMs (KR/US/CN) on cybersecurity: closed-book knowledge (WMDP-cyber, CTIBench) vs agentic CTF-solving (Cybench)

agent-attack-atlas：面向工具型 LLM 智能体的攻击方法图鉴与实测：188 张方法卡+ 13 条确定性判据 + 36 批次 1664 条可离线复算的证据束。

Benchmarking how well agents use embodied skills on physical tasks.

A data-incident agent whose fixes must pass an authorizer and an independent validator, and the evaluation behind the paper "Permission Is Not Justification": a frozen twin benchmark, registered runs, and a checker that recomputes every reported result from the published evidence.

Proof of Agent Work (PoAW): an open protocol and self-hostable node that verifies what AI agents claim they did and issues signed, publicly checkable receipts.

Claude Code plugin: a knowledge base of your project's conversations and code that builds and updates itself (MemPalace + Graphify)

Whole-repository code review by AI agents: blocks, a coverage map, hypotheses with verdicts, four roles. An Agent Skill.

AI claims triage for car damage: custom-trained vision models as tools for an auditable LLM agent. Models measure, the agent reasons, rules decide.

Open-source MIT-licensed multi-agent LLM-driven simulation of the AI startup ecosystem — a what-if scenario tool for investors, VCs, and strategists

Agent skills that build Slack and Discord bot harnesses — collaborating characters with narrow, code-enforced authority.

The execution layer for AI tax prep — agents think, taxfill fills, verifies, and gets it mailed. Guided intake + deterministic fill/verify/calc tools + versioned form & jurisdiction knowledge, for any MCP client.

ARIA: Ask Your Advisor - Learning when to consult stronger models in agentic systems

Claude Code plugins for reviewing changes: a four-pass code review, migration safety, regression hunting from git history, change-impact maps, tests from findings, and an orchestrator that runs what a change needs.

Multi-agent AI research pipeline: researches a problem area on the web, finds unsolved gaps, stress-tests them with a critic and majority vote, and designs solutions checked against existing work.

Manage your job search like a software project through GitHub with AI automation. Jobs are tracked as GitHub issues and projects. Agents scrape, triage, research open roles and update application status.

Evolution for the Agent Era. Inspired by Darwin, an open framework for experience-driven recursive self-improvement of AI agents.

Claude Code skill that delegates large reads and code writing to local models (Ollama) to save Claude tokens - with automatic checks, sandboxing and escalation. Never the cloud.

A curated list of environments, benchmarks, datasets, and RL training resources for workplace AI agents: email, chat, calendar, docs, and enterprise apps.

Spec-driven software-engineering team of three AI roles (Morfeo, Supervisor, Implementers) on Hermes Agent: Objective Contracts, isolated worktrees, independent review. v1.0.

AI Work Continuity & Cross-Model Agent Handoff System. Captures, verifies, and packages deterministic project state across Claude, GPT-4o, Gemini, Cursor, Antigravity & Local LLMs.

Pool multiple free-tier LLM API providers behind one OpenAI-compatible endpoint, with automatic failover and a live dashboard.

用自然语言查询经营数据的受控 Agent：模型只提议查什么，服务端核权限、执行只读 SQL、核实数字，回答里写明口径和依据。

Open-source AI cybersecurity companion

Gymnasium environment where AI agents solve SPaRC grid puzzles one move per step, with optional backtracking. Comes as a pip package, a Prime Intellect verifiers env and an OpenEnv server. The best model solves 16.0%, humans 98.0%.

FastAPI + LangGraph AI Agent backend with RAG, JWT auth, SSE streaming

No Code Multi Agent Orchestrator. Open source alternative to your favourites n8n, Gumloop, Relevance AI etc.

基于 RAG（Retrieval-Augmented Generation） 技术的服装领域知识管理系统。将服装搭配、洗涤养护、尺码推荐等知识文档向量化存储，支持通过 Web 界面上传新知识并自动去重。

OpenTelemetry-native agent observability: trace every LLM call, tool call and agent step, with cost per call, alerts, and Claude Code / Codex / Cursor session tracing. Python SDK.

static region:us

Did this change to the coding-agent harness make the code better or worse? Paired, repeated, isolated evaluation with a verdict you can trust.

A time machine for AI agents.

ZhiXi LingXi — 24/7 full-stack digital employee engine

Sandbox for comparing Agentic & RAG frameworks

Agentic-Harness-Engineering-Workshop

Self-hosted AI code reviewer for GitHub pull requests. DeepSeek-V4 + Blackgeorge multi-agent pipeline posts diff-anchored findings with severity badges and suggestions. Runs as a GitHub Action or your own GitHub App. BYOK.

English-first, Traditional Chinese companion: multi-agent coordination handbook, diagrams and safe starter templates for Orca.

Multi-agent AI system for autonomous retail pricing optimization using Amazon Bedrock AgentCore. 6 specialized agents orchestrate competitive intelligence, demand forecasting, and strategy synthesis to generate ranked pricing scenarios with human-in-the-loop governance.

AI 에이전트로 전문 투자 리서치 팀을 구성하는 Claude Code 프레임워크 — 병렬 서브에이전트 4관점 분석 + 결정론적 검증 도구 + 워크포워드 백테스트

一个人指挥一群 AI 窗口干活时用的一套 md 文件，外加一个工单台。

Local-first team workbench for AI agents: project mailboxes, shared context, task handoffs and human review. v0.8.2 · macOS App and local Web; Intel, Windows and Linux previews.

Track health data, detect patterns, research science, and design self-experiments with AI agents running on Claude Code and OpenClaw.

An open-source engineering operating model for Codex: model-aware subagents, bounded delegation, task graphs, independent review, and validation.

Route Codex tasks between GPT-6 Astra and GPT-5.6 Sol by expected rework cost, then delegate bounded repetitive work to GPT-5.6 Luna.

本地优先、透明可见无黑盒的多项目、多智能体 AI 工作台 / A local-first, transparent workspace for multi-project and multi-agent AI collaboration.

Run Claude Code, Codex, Antigravity, Ollama and more side by side — group chats where AIs work together, controlled from your browser, Telegram or WhatsApp. Free & open source.

Human-in-the-loop control surfaces for monitoring and directing AI agents across Novation Launchpads, browsers, phones, and tablets.

Self-hosted memory for LLM agents - one SQLite file, zero services, MIT. Hybrid retrieval (FTS5 + sqlite-vec + RRF) with pain-weighted salience and a knowledge graph. Under matched embeddings, outperforms Mem0 on LoCoMo and LongMemEval. Your data, your provider, your disk.

Give your AI team a mission, authority, and accountability. Four Markdown skills for delegated decisions, coordinated execution, independent QA, and checkpoint learning. Codex-native; adaptable to Claude Code and OpenCode.

AI grant-writing agent on LangChain Deep Agents: finds funding opportunities, fit-scores them with quoted evidence, and drafts proposals with human approval of final files. Marks unknowns for input instead of inventing data.

Swarmattacker Web application scoped Black box penetration testing agent

A persistent AI agent with its own life, not an assistant. Wakes on its own, keeps its own memory, built by hand one piece at a time.

Autonomous coding agent built on LangGraph + deepagents. It natively plans, builds, verifies, and ships real production changes with a built-in human-in-the-loop review stage.

ShoppingClaw(虾购)是采用Langgraph架构设计的Mutil-agent电商智能消费决策系统。比参数、比价、看口碑、防坑——买一件东西比上班还累。ShoppingClaw 把这条链路交给多个 Agent 自动化：搜货、比价、取舍、风险排查、归档、盯价。一句话提需求，推理过程实时铺开，可随时打断。

Open Model Hardware Protocol: an MCP-style open protocol for AI agents to find, understand and safely operate lab and factory hardware

Embeddable self-adaptive agent kernel for host applications

DeepTrace: a fault-tolerant deep research agent that traces every claim to its source. Claim-level faithfulness checks, crash-safe tool calls, leased multi-worker execution, FastAPI + Docker.

Python agent framework where the agent loop is a function you write

Drop-in plugin that brings plan-first, human-supervisable, traceable, and parallel agentic coding to many AI harnesses.

Claude Code skills that make sessions resumable and repos self-documenting: plans, idea backlog, docs capture, PR review, session wrap-up. Installable plugin marketplace.

Project template for AI coding assistance: rules, skills, subagent roles and enforcing hooks, built primarily for Claude Code.

Run process-correctness validation as a custom code-based evaluator in Amazon Bedrock AgentCore Evaluations, deterministic behavioral contracts over the OTEL traces agents already emit, with audit evidence records.

Best blend of spec-driven and superpowers - YMMV

1min.AI — Business SaaS listing. Visit link via SOFTGIT. Third-party; rights belong to original authors.

Self-hosted personal AI assistant and Muse alternative with cloud code/browser sandbox, MCP and web/desktop/iOS clients. Source-available; personal noncommercial use is free, commercial use requires written authorization.

Your body's data layer for AI: a self-hosted scaffold any AI (Claude, ChatGPT, Codex, DeepSeek…) reads and writes via MCP. Apple Watch data, WHOOP-style scores, no forms. Not affiliated with WHOOP or Apple.

J.A.R.V.I.S — An advanced AI voice assistant for macOS with Hindi/English support, powered by Llama 3 70B and Groq.

🧸 Put your Muse in Discord. DM it or invite it to any Discord server.

Bypass Huawei Health lock-in. Watch GT 5 data → Gadgetbridge → Termux → AI analysis → WeChat push. Open-source health data pipeline.

Windows 微信 AI 助手：多会话人设、长期记忆、本地维护台与高级设置。Clone and reproduce with your AI.

栗栗 (Tamias) — an AI agent desktop pet & virtual assistant for Windows. Squirrel detective that chats and does real work (reads files, writes code, edits files) with one-click rollback. 替你干活、改坏能回滚的桌面 AI 管家

A macOS desktop app that listens along to lectures and meetings: live speech-to-text, timestamped answers about what you missed, and editable notes next to the transcript

A computer-use agent with its own desktop and its own cursor — Windows 11 and Linux. It works while you keep using yours. Claude computer-use + PySide6.

Personal voice AI agent for macOS: real-time voice, persistent context, and plugins for Gmail, GitHub, and Notion.

Your own JARVIS for Linux: a voice-and-keyboard AI agent for KDE Plasma 6 that does things — apps, windows, music, mail, images, video. Claude or free models, speech stays on your PC · Личный Джарвис для Linux

A Hyprland desktop based on Arch Linux: living wallpapers that react to your machine, panels with real readings, holograms, your homelab on screen, and VECTOR, an AI with a voice.

Your AI assistant on the Mac menu bar, iPhone and Apple Watch. A thin client of AbstractGateway: talk or type, hands-free voice conversations, automations, and sessions that follow you across devices. Any provider and model, local or cloud.

Local AI assistant: one Qwen3.5 model plus a library of LoRA experts, picked for each message and checked in code. Runs on a 16 GB laptop, OpenAI-compatible API.