Waterline is a free macOS screensaver by Ascenda that visualizes AI coding agent activity through rising water and ripples, supporting Claude Code, Codex, and Cursor. It runs locally without network connections or account requirements, with optional Ascenda Flow for enhanced tracking and color representation of workload.
Eval-skills is a framework for building and improving AI agent applications through systematic evaluation and experimentation. It guides developers from inspecting real executions through human review, error analysis, and validated grading to running controlled improvement experiments with Descent. The tool integrates with Claude Code, Codex, and Cursor to help teams establish trustworthy baselines and catch regressions in production.
Claude Code and Codex agents exhibit incompatible MCP server support, with five of ten tested features failing on one agent while working on the other. The Harness Quirks Matrix, built on the M3 testing tool, identified specific failure modes: Claude Code drops tools with ids[] properties and truncates error messages, while Codex caches outdated tool catalogs and compacts large schemas. Eight additional failures were found through direct binary probing.
Zulu Assistant is a Mac-based ops board and agent workforce platform that monitors LaunchAgents, containers, cron jobs, and endpoints while managing Claude Code and Codex sessions. It provides visibility into background job failures, agent session status, and coordination with human approval gates for risky operations, running as a zero-dependency Node 22 server with policy-driven escalation.
Faplex is a multiplexer for managing Claude agents, Codex, and OpenCode sessions across multiple machines via SSH, displaying a unified real-time status list without requiring setup on remote harnesses or persistent terminal panes.
Clanking is a tool that converts LLM coding agent activity into audible soundscapes, using different tones and positions to indicate agent states like working, waiting, or errors. It runs as a Unix daemon with PipeWire, synthesizes sounds quantized to musical notes, and integrates with Claude and Codex through plugins.
A solo developer has created an unofficial VR mod of the 2005 horror game F.E.A.R., bringing the original campaign to PC virtual reality with tracked hand controls and physical weapon handling. The free mod, developed with assistance from OpenAI's Codex AI tool, requires a legitimate copy of the base game and a compatible VR headset.
SpecWeave 3 is a CLI tool that enables handoff of coding tasks between multiple AI agents (Claude Code, Codex, Grok) while maintaining context and work state. It uses spec-driven planning with acceptance criteria, automatic handoffs near usage limits, and generates audit trails through HTML reports.
Wy is an experimental Rust terminal tool that connects Git diffs to AI-generated code explanations from Codex or Claude Code conversations, allowing developers to review code changes alongside the agent's reasoning and inspect cited evidence offline.
Clodex is a fork of OpenAI's Codex that replaces the backend model with Claude, integrating Anthropic's Messages API directly into the ChatGPT desktop app without proxy or translation layers. Built in a single day, it maintains the existing ChatGPT UI and voice mode while routing requests to Claude, requiring both ChatGPT and Claude Code logins.
ModelRudder adds experimental per-task model routing to native Codex and Claude Code, allowing users to allocate stronger models to tasks that need them. The free, MIT-licensed preview runs locally on macOS, Linux/WSL, and Windows, requiring Node.js 24+ and a TypeSafe Jev API key for routing decisions across Haiku, Sonnet, and Opus models.
OpenAI launched Ultrafast mode for GPT-6.1 Sol today, offering speeds up to 8x faster than standard mode with intelligence approaching GPT-6 Astra. The feature is rolling out across API, Codex, and ChatGPT Work, with API pricing at $12 per million input tokens and $60 per million output tokens.
A developer describes configuring Codex into a multi-agent system with specialized sub-agents bound to different models to reduce token consumption and improve context window efficiency. Rather than using a single powerful model for all tasks, the approach routes different job categories to appropriately-sized models, saving tokens while maintaining output quality.