Open Source Agent Tools: November 5, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. AutoCua 0.1.8 and 0.1.9 Run Coder, Browser and Phone Sub-Agents in Parallel and Close a Windows TLS Hole
AutoCua, a Python and Rust computer-use agent by Ashish Yadav with about 1,250 GitHub stars, shipped v0.1.8 on November 3 and v0.1.9 on November 4 and was posted to Show HN on November 4. A parent agent works macOS, Windows or Linux desktop apps by choosing numbered elements from a view that fuses the OS accessibility tree with OCR (the model never sees coordinates, and an id it was not shown is refused), sends input at the OS input layer, and delegates to a coder agent with 14 tools and read-only “minion” scouts, a browser agent built as one Rust crate that drives the user’s own logged-in Chrome through an extension with hand-written CDP (no Playwright or Selenium), and iOS and Android agents over USB or Wi-Fi. Version 0.1.8 adds a Daisy chain mode that brings a paired iPhone and Android phone in as sub-agents, plus fixed slots for up to 4 coder agents and 4 browser agents at once, each browser agent in its own background tab of one Chrome. Version 0.1.9 fixes a Windows flaw in versions 0.1.0 to 0.1.8 (GHSA-qm8g-cw9j-w2gj) in which a failed certificate check switched HTTPS verification off for every connection, letting an attacker on the same network read API keys, the Telegram bot token, prompts and screenshots; it now falls back to the Windows certificate store and otherwise stops. It supports 60 models across 10 providers including Anthropic, OpenAI, Google, AWS Bedrock, Groq and OpenRouter, a separate agent compresses the transcript mid-run so long tasks do not hit the context limit, vault credentials are typed by the runtime without being shown to the model, and the project says it sends no telemetry; it installs with pip install autocua (Python 3.10 or later) as wheels for Apple Silicon macOS, Windows x64 and Linux x86_64, and iPhone control requires a Mac. License: MIT. Source v0.1.8 Repository Show HN
2. Screenplay Lets Claude Code, Codex or Cursor Script, Film and Edit Screen Recordings Over MCP
Screenplay (agent-screenplay on npm), created on November 2, released through v0.1.7 on November 3 and posted to Show HN on November 4, is an MCP server, CLI and skill with which a coding agent explores a website or desktop app, writes a JSON scenario of steps (clicks, typing, callouts, spotlights, zooms, captions and optional title cards), and has Screenplay play it, film it and edit it in the style of Screen Studio, with a smoothed cursor, automatic zooms that follow the action and click effects. Screenplay itself contains no AI and needs no model key: npx agent-screenplay init downloads Chromium once, registers the MCP server with any Claude Code, Codex or Cursor install it finds and adds the scenario guide, and Claude Code gets a /screenplay command. Web runs use Playwright with headless Chromium, while desktop apps on macOS, Windows and Linux X11 (Wayland is not supported) are driven through accessibility APIs by Vercel’s agent-desktop and filmed with ffmpeg, which needs a real screen. Elements are targeted the way people see them (text=, role=button[name=...], label=), check rehearses the steps without filming, render re-edits the last take without filming again, and voice-over can use a Cartesia or ElevenLabs key or the user’s own audio files. Output defaults to 1920x1080 at 30 fps, up to 4K and 60 fps or a GIF; the README warns that steps really run (a Send click sends) and recommends demo accounts or staging. License: MIT. Source Show HN
3. Amigent Gives Coding Agents an MCP Server and Remote Desktop for a Real Commodore Amiga
Amigent, posted to Show HN on November 4 after its v1.1.1 tag on November 2, connects a PC or Mac to a real Amiga on the home network so an agent such as Claude Code or Codex can install and update software, change settings, copy files and click through programs over MCP, while the user watches every click in a remote-desktop window and can take over at any moment. A small amigentd daemon runs on AmigaOS 2.04 or later on any CPU from the 68000, needing a TCP/IP stack such as Roadshow and a network link as slow as a plipbox on the parallel port; the host side installs with pipx install amigent on Windows, macOS or Linux with Python 3.10 or later, and the README suggests letting the agent run the setup itself. Screens travel as the Amiga stores them (bitplanes, or only the changed palette tiles), and input is handled before pending picture data; the author measures 0.6 to 1.0 seconds for a full 640x256 Workbench screen on a 7 MHz 68000, and 1.9 seconds for a first 1920x1080 picture over a roughly 50 KB/s plipbox versus 124 seconds as raw pixels. A key keeps other machines on the network out, the connection can be encrypted, and the author warns against exposing the Amiga to the internet. License: MIT. Source Repository