harness-engineering

infiniflow/ragflow

RAGFlow pairs deep document understanding with agentic RAG — RAGFlow combines retrieval-augmented generation with agent capabilities, grounding answers in traceable citations from complex documents. It targets teams building LLM apps that need reliable, well-cited knowledge bases rather than simple chatbots, with multi-language docs and a cloud-hosted option.

agentscope-ai/QwenPaw

QwenPaw: a self-hosted personal AI assistant with layered memory — QwenPaw is a self-hostable personal AI assistant built on Alibaba's agentscope stack, deployable locally or in the cloud with no mandatory API key via its built-in QwenPaw-Flash models or Ollama/LM Studio. It adds a three-layer memory system (working context, full history, self-evolving knowledge base), sandboxed security guards, and multi-agent orchestration, aimed at users wanting a private, extensible assistant across chat channels.

langchain-ai/deepagents

A ready-to-run agent harness built on LangGraph — Deep Agents is an opinionated, batteries-included agent harness from the LangChain team, aimed at developers building long-horizon, multi-step AI agents. It ships sub-agents with isolated context, a pluggable filesystem backend, and automatic context summarization, while staying model-agnostic and overridable piece by piece.

pydantic/pydantic-ai

Pydantic's typed agent loop covers voice, images, and every LLM — Built by the Pydantic team, this SDK offers a typed, extensible agent loop that works with any model via a simple string swap. The same agent runs across web, CLI, voice calls, and durable queues, aimed at Python developers wanting type safety across the whole AI stack.

HarnessMD/munder-difflin

An open-source office of coding-agent clones on your laptop — Munder Difflin is a free, local-first multi-agent harness that wraps Claude Code, Codex, Gemini CLI, Grok, OpenCode, Cursor and more, using your existing subscriptions. It visualizes agents as avatars on a shared office floor, coordinated by a 'clone' that routes messages and remembers context while you're away.

HarnessRouter/harnessrouter

One API to run Codex, Claude Code, Hermes and more agent harnesses — HarnessRouter CE is a self-hosted, Apache-2.0 gateway that unifies agent harnesses like Codex, Claude Code, Hermes, PI and DSH behind a single API with sessions, streaming, files, cancellation and failure handling. It implements the open Unified Harness Protocol (UHP) and is OpenAI Responses-compatible, aimed at teams building agent products who want to avoid harness-specific integration work.

openbkn-ai/bkn-foundry

Ontology backend that turns business knowledge into executable services — BKN Foundry is the Go-based backend of OpenBKN, converting ontology models of data, rules, and processes into runtime services with data access, safe execution, and governance. It's a backend-only framework (no UI), aimed at developers building agents that need traceable, verifiable business logic rather than just generated answers.

HelpCode-ai/anythingmcp

Self-hosted gateway that wraps 517 APIs as MCP tools — AnythingMCP connects REST/OpenAPI, SOAP, GraphQL, OData and SQL systems to Claude, ChatGPT and Copilot without writing a custom MCP server. It ships 517 ready-made connectors, including SAP, Etsy and Amazon Seller, with 23 needing no API key, plus RBAC, SSO and audit logs for enterprise use.

Agent-Field/CodeAF

A coding harness built to squeeze frontier results from open models — CodeAF is a single-binary coding harness tuned for open models like DeepSeek, Qwen, GLM, Kimi and MiniMax rather than closed APIs. It replaces juggling multiple terminal agents with one window that tracks work across projects, letting you hand off tasks and step in only when needed. Early preview, Apache 2.0, runs on macOS/Linux/Windows.

uclanlp/OpenVLHarness

Official Code for OpenVLHarness: Real-World Visual Agentic Tool‑Use with Multimodal Memory

rudderlabs/devbox

Open source autonomous engineering harness in your cloud. One sandbox per task, no real secrets inside.

sshah03/recalc

A spreadsheet grid where every cell is its own AI agent terminal — recalc reimagines the spreadsheet for agentic AI: each cell can hold a value, formula, terminal command, or a Claude Code/Codex/pi agent. You watch agents fill cells live, open any cell as its own terminal, and steer work by prompting or editing cells directly. It's an early-beta Tauri desktop app for developers orchestrating multiple coding agents at once.