Code does code work. AI does AI work.
Open-source tooling for coding agents — Claude Code, Codex, OpenCode, Cursor, Kiro, and MCP hosts.
24 plugins · 50 agents · 45 skills · 423 lint rules · 15 desktop MCP tools
New plugins, validators, and MCPs ship constantly. Follow for updates:
| Project | What it is |
|---|---|
| agentsys | Plugin marketplace and orchestration runtime for coding agents. 24 plugins, 50 agents, 45 skills across Claude Code, Codex, OpenCode, Cursor, and Kiro. |
| agnix | Linter and LSP for agent configuration. 423 rules across CLAUDE.md, AGENTS.md, SKILL.md, hooks, MCP — the validation skills don't get anywhere else. CLI, LSP, editor extensions, GitHub Action. |
| computer-use-linux | MCP server for native Linux desktop control. AT-SPI accessibility trees, real Wayland support, multi-compositor window targeting (GNOME, KDE, Hyprland, i3, COSMIC). 15 tools, Rust. |
| agent-workspace-linux | Isolated Linux desktops for AI agents. A hidden, agent-owned desktop and browser over MCP, so an agent does GUI and web work without touching your real desktop. Rust. |
| parlar | Voice conversation mode for Claude Code and Codex. Talk to a running session and it answers out loud while it works; push through hooks, no terminal injection. Local CPU models (Phonon-2, Kokoro), GNOME swarm indicator. Rust. |
npm install -g agentsys && agentsys
npx agnix .
npm install -g @agent-sh/computer-use-linux
cargo install parlar parlard && parlard fetch && parlard serviceAI writes code. That's the easy part now. The work around it was the bottleneck:
- Skills don't trigger when their config has the wrong field. Vercel's research found skills invoke at 0% without correct syntax — and you find out after the fact.
- "Computer use" MCPs only work on macOS. The Linux options drive
xdotoolagainst X11 and pretend Wayland doesn't exist. - Context doesn't survive a session. You restate the same project decisions every time, or you bloat
AGENTS.mduntil it stops being read. - Reviews happen by vibes. CI breaks at the worst time.
I built these tools for myself first. Posting them here in case they're useful to you too.
Three principles shape everything in this org:
- Deterministic tools do detection. Models do judgment.
- A workflow is a pipeline with gates, not a monolith.
- Configuration should fail loudly before production work depends on it.
Also in the ecosystem — plugins, memory, infra
Workflows · next-task · prepare-delivery · audit-project · ship · gate-and-ship · deslop · enhance · sync-docs · drift-detect · perf
Memory & curation · axiom · banthis · skill-curator · system-prompt-curator
Skills · onboard · can-i-help · learn · consult · debate · skillers
Infra · agent-analyzer · agent-core · repo-intel
Bridges · glidemq
Built and maintained by Avi Fenesh — Valkey contributor and Valkey GLIDE maintainer. Solo-maintained. Contributions welcome if they fit — please follow CONTRIBUTING.md.