Open Source Agent Tools: November 22, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Pi Bot Is an Apache-2.0, Local-First Desktop App Where a Team of Named Agents Works in Group Chats With Approval Gates
Pi Bot runs named Bots on your own computer, each with its own role, model, skills, permissions and memory. A group chat has a coordinator that plans and hands steps to members, and the user confirms the plan before it runs. Every tool call passes one policy engine, commands run in a sandbox built on @anthropic-ai/sandbox-runtime, and approvals are written to an audit log. It supports Ollama, OpenRouter and your own keys, MCP servers and skills, runs on macOS, Windows and Linux with an Android app, and is built on the pi agent kernel. License: Apache-2.0. Source
2. Anvil Is an AGPL-3.0 Native macOS Coding Agent That Works With Almost Any Model Through a Plain-Text Tool Protocol
Anvil is a SwiftUI desktop coding agent in the spirit of Claude Code and Codex, at version 0.1.0 and requiring macOS 15. Its agent speaks a plain-text protocol instead of native tool calling, so it runs on any OpenAI-compatible chat model, and it adds its own checks: tests run before “done”, a reviewer reads the diff, and a different model takes over after repeated failures. Every edit is a diff before it is written and every request is a checkpoint that can be undone. The README reports 16 of 16 passing evals. License: AGPL-3.0. Source
3. Nebula Memory Is an MIT-Licensed, Zero-Dependency Local Memory Engine for Agents With Four-Way Recall and a Visual Dashboard
Nebula Memory is a self-contained agent memory system written in Node.js that stores memories in local JSONL and needs no database, vector store or LLM. Recall combines four paths (lexical, activation, graph diffusion and profile), an interactive dashboard visualizes the memory, and an optional Hindsight HTTP API adds semantic matching. It exposes an MCP interface and positions itself against heavier memory frameworks such as Mem0 and Zep. The README is mostly in Chinese. License: MIT. Source