Open Source Agent Tools: November 8, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Ryph Is a Self-Hosted Orchestrator Whose Coding Agents Pick Up Alerts, Work in the Incident Chat and Return Hotfix Pull Requests
Ryph was posted to Show HN on November 7, after its first public release (v0.1.0) on November 4 and three fix releases (v0.1.1 to v0.1.3) on November 6. It is a self-hosted orchestrator for team coding agents: a run starts from an alert webhook (Grafana, Alertmanager, Grafana IRM, PagerDuty, Sentry, incident.io, Datadog, Opsgenie or a custom mapping), a task in the UI, a /ryph comment on a pull request, merge request or issue, a chat mention or a cron schedule. Each run is one container with one agent that holds only credentials minted for it and reaches models and tools through Ryph’s model proxy and MCP gateway, so real keys never enter the container. For incidents, Ryph opens a “response” bound to one Slack thread, Zulip topic or Telegram group; later alerts about the same problem join it as messages, people steer the agent by mentioning it, a memory package supplies runbooks and gotchas left by earlier runs, and the forge layer (GitHub, GitLab, Radicle or a local directory) turns the agent’s patch into a branch and pull request after a person approves it, or immediately where a repository allows that. Models can be Anthropic, OpenAI, a ChatGPT plan or any OpenAI-compatible endpoint, and runs can also use repo-declared checks in .ryph/checks.yml, dev-server previews, an in-sandbox Chromium and MCP servers with a per-agent policy. It ships as a single binary with embedded PostgreSQL, a Docker Compose setup or a Helm chart, and the README warns that breaking changes are possible before v1. License: MIT. Source Release Show HN
2. Conducted Rebuilds Conductor’s Local Parallel-Agent Workspace as an MIT-Licensed Mac App
Conducted was posted to Show HN on November 7 with v0.3.1, its first published 0.3 release (v0.3.0 was tagged but its build failed), followed by v0.3.3 on November 8. It is an independent open-source rebuild of the local features of Conductor, the app for running coding agents in parallel whose team announced on November 4 that it is joining Meta. Conducted drives Claude Code, Codex, Cursor, OpenCode and Pi under the user’s own subscriptions and keys: each task gets its own git worktree and branch, copied .env files, a setup script and a reserved port; each workspace can hold several chats with any agent and model, with queued messages, mid-turn steering and plan approval cards; and a hidden git ref before and after every turn lets the user revert to or fork from any turn. Review covers diffs by branch, commit or turn, with line comments sent back to the agent, and the same panel creates pull requests, watches checks, returns failing checks to the agent and merges. Team features are left out, and cloud workspaces are being built to run on the user’s own Cloudflare account (v0.3.2 added SSH into a cloud workspace and opening it in VS Code, Cursor or Zed). The maker says Conductor’s behavior was documented and rebuilt from that description without copying code, that Conductor 0.92.1’s own frontend runs on Conducted’s Rust shell and sidecar, and that the roughly 197,000 lines of TypeScript and Rust were written by DeepSeek flash agents orchestrated by Opus 5.5 with Sonnet reviewing each batch, for about $50 of DeepSeek usage plus about 70% of a week of a Claude 20x subscription. It needs an Apple silicon Mac and installs from a signed DMG or a Homebrew cask. License: MIT. Source Release Show HN
3. backendskills Reports Critical Backend Mistakes Falling From 25 to 2 in 150 Agent Runs Scored by Hidden Checks
backendskills, posted to Show HN on November 7 (the repository was created November 5), packages three agent skills for Claude Code, Codex and Cursor: api-endpoint (scope every query to the caller, turn bad input into a 4xx, rate-limit login and costly routes), data-integrity (locks or conditional updates for read-then-write, UNIQUE constraints, idempotency keys on anything that charges or sends) and db-migration (expand then contract, concurrent indexes, lock timeouts, batched backfills, down migrations). The authors gave agents 15 backend tasks with and without the skills and scored them with hidden checks the agent never sees, such as 20 buyers at once for 5 items in stock, a user reading someone else’s order, and the old app version running against the new schema mid-deploy. With Sonnet 5.5 and Haiku 5.5 at five runs per task, runs with a critical mistake fell from 25 of 150 to 2 of 150 (Haiku 12 to 2, Sonnet 13 to 0); 20 of the 25 failures came from login rate limiting and column renames, which failed in every run without the skills and in none with them. Missed team conventions such as down migrations and API docs fell from 30 to 4, while Haiku still mishandles out-of-order payment webhook events (1 of 5 runs failing without the skills, 2 of 5 with them). The tasks and every raw run are in the repository’s bench/ directory, so the numbers are the authors’ own but reproducible. It installs with npx skills add globalmpc/backendskills, or as dist/AGENTS.md for Codex and other AGENTS.md agents and dist/cursor/ for Cursor rules. License: MIT. Source Show HN
4. ywza Lets a Model Operate a Windows or X11 Linux PC From Paired Devices, Reading Windows as Numbered Element Lists
ywza, created and posted to Show HN on November 7, is a computer-use agent written in Rust: ywza-agent runs in the background on the PC to be controlled and can look at windows, click and type in any app and work in open terminals, while the ywza CLI sends it instructions from the same machine or another one. Devices pair once with a 6-digit code that works once and expires after 10 minutes, then connect peer to peer through public relays (running your own relay is planned), with no account or hosted server. The model is bring-your-own: anything that speaks the OpenAI chat API, including local Ollama and LM Studio. To keep token counts low, ywza describes each window to the model as a numbered list of elements rather than sending a screenshot at every step, though the model can still request one, and it can be built with Windows OCR or PaddleOCR (through ONNX Runtime) for apps that expose nothing to accessibility APIs. An MCP server gives any MCP client control of the local PC and every paired device, with the model choosing the device per call. It targets Windows, supports Linux only under X11 (with a systemd unit for the login screen and user sessions) and builds from source. License: AGPL-3.0. Source Show HN