Open Source Agent Tools: November 20, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Mingle Is an MIT-Licensed Local App That Puts One Coding-Agent Companion in Slack, Telegram, Discord, Feishu and WeChat
Mingle (earayu/Mingle) creates an agent on the user’s own computer, with a name, a role, a memory and a home directory, and lets an existing coding agent such as Claude Code, Codex, OpenCode or Pi drive it. The same agent joins the user’s chat apps as a bot, including WeChat, Feishu, Slack, Discord, Telegram, GitHub and iMessage, keeps one memory across them, works on long tasks in the background while still answering others, and can be stopped by the owner with /stop. There is no Mingle server, account or telemetry, and it runs on the harness subscription the user already has. The repository is MIT licensed, was created on November 11, and shipped v0.4.4 on November 20. Source
2. Agentest Runs 165 Offline Checks on an Agent’s OpenTelemetry Traces to Catch Claimed Success It Did Not Earn
Agentest (srpadmanabhuni/agentest) is an Apache-2.0 evaluation tool that reads the OpenTelemetry GenAI traces an agent already recorded and judges what the agent did rather than what it said, such as claiming success without the matching tool call, acting before verifying, following a hidden instruction, calling a forbidden tool or leaking personal data. It runs offline with no API key, network or LLM judge, so the same traces always give the same verdicts. It installs with pip install agentest-ai and has native-log importers for LangGraph, the OpenAI Agents SDK, AutoGen, smolagents and n8n. It was a Show HN on November 19, the day v3.5.0 and v3.5.1 were released. Source
3. Dynobox Is an Apache-2.0 Test Runner That Asserts on Tool Calls, Files and MCP Requests Across Six Coding-Agent Harnesses
Dynobox (dynobox/dynobox) runs real coding-agent workflows in fresh temporary directories, captures evidence, and asserts on tool calls, commands, files, HTTP requests, transcripts and final answers without one model judging another. The same scenario can run through Claude Code, OpenAI Codex, OpenCode, Pi, Cursor CLI and Google Antigravity CLI to expose behavior that differs between harnesses, and runs can be repeated to find flaky behavior. It also offers experimental CLI mocks and HTTP MCP tool fixtures, requires Node.js 22 or later, installs through npx dynobox@latest init, and was a Show HN on November 19. Source
4. Bot-inator Is an MIT-Licensed Local macOS Agent That Proposes Its Own Tools, Then Runs Them in a Sandbox After Review
Bot-inator (Bforsyth1234/bot-inator) is a menu-bar macOS agent that watches activity such as new files in Downloads, mail, iMessages, meetings and app switches, asks a model how it could help, and shows each offer as a card; nothing that changes state runs without approval. When it spots a repeated manual task it can offer to write a tool, checks the code statically and with a second model pass and a throwaway sandbox dry run, shows the full source for approval, and then runs each call in a short-lived child process under macOS sandbox-exec, with tools versioned in a git repository. Models run locally on MLX by default and cloud providers are opt-in. It requires Apple Silicon and macOS 14, and v0.2.0 through v0.2.2 were released between November 19 and 20. Source