fast-jev-compaction leads at 1759 stars/day
We don't tell you what's popular — popularity lags and is gameable. We tell you what's gaining speed right now, and whether it's worth your attention. 6 accelerating AI repos earned today's slot.
Top mover
A Claude Code plugin that swaps the default compaction summary for a per-tool-call scoring pass: each tool call and result gets a fast relevance score, stale entries are dropped or truncated, and retained content stays verbatim rather than paraphrased. This targets a real pain point — lossy summarization in long agent sessions — and the verbatim-retention angle is a meaningful design choice over summary-based context management. Caveat: it depends on an extra scoring request per compaction cycle, so latency and token cost scale with session length, and the scoring heuristic's quality is unproven at 3.5k stars.
🤖 Agents & automation
fx is a terminal coding agent built around Unix composability: it pipes file context and shell output through an LLM loop rather than locking you into an IDE or chat UI. It's gaining traction because it's from Vercel Labs and slots into existing shell workflows, but it's early and overlaps heavily with Claude Code, Aider, and OpenAI's Codex CLI.
grok-build is a terminal coding agent harness with a fullscreen, mouse-interactive TUI and an extensible plugin surface, shipped by xAI and pulling ~27k stars fast on brand gravity plus genuine demand for agentic CLIs. It matters because it competes directly with Claude Code, Aider, and Codex CLI on ergonomics rather than raw model access, and the extensibility hook is the differentiator. Caveat: it's tightly coupled to xAI's Grok models and API, so portability and long-term maintenance are unproven.
Tooling & infra
zvec-grep combines BM25 full-text search with vector/semantic search over a local workspace, exposing results via CLI and an MCP server so both humans and AI agents can query the same index. It's gaining traction because agentic coding workflows need fast, private, offline retrieval that doesn't ship source code to a hosted API, and ripgrep alone can't do semantic matching. The caveat: hybrid search quality depends heavily on the embedding model and index freshness, and maintaining a vector index over a large monorepo adds real disk and reindexing overhead versus plain grep.
Codex-X wraps the OpenAI Codex desktop/CLI client with a visual layer for switching providers and API keys, syncing sessions, injecting prompts, and managing Skills/MCP servers plus TOML config. It is gaining traction because Codex users otherwise hand-edit TOML and juggle multiple API endpoints, and this centralizes that workflow cross-platform. The caveat: it is a third-party front-end over a fast-moving OpenAI surface, so it can lag or break when Codex's config schema or auth changes.
Codex-Dream-Skin is a community theming/customization layer for OpenAI's Codex CLI, letting users restyle the terminal agent's UI and prompts. It's gaining traction because Codex CLI adoption is spiking and developers want personalization and workflow tweaks the official tool doesn't ship. The caveat: it's a cosmetic/UX wrapper, so its value depends entirely on upstream Codex CLI stability and it adds a maintenance dependency for little functional gain.