One coding agent became a fleet, and the tooling to herd it is arriving.

A terminal multiplexer built for coding agents: a ~10MB Rust binary that gives each agent its own real terminal (so full-screen TUIs render right) plus a status board showing who's blocked, working, done, or idle. It detects Claude Code, Codex, Cursor, Copilot CLI, Devin, Grok CLI, Droid and ~10 others, and will multiplex any custom agent like a normal terminal multiplexer. Install is curl -fsSL https://herdr.dev/install.sh | sh (also Homebrew and Nix; Windows in beta). Persistent sessions survive a disconnect, there's a socket API for orchestration, and there's no GUI, Electron, account, or telemetry.

Reach for it the moment you're alt-tabbing between three or four terminal windows to find which agent is sitting blocked on a prompt. Delete the tmux-with-extra-config setup or the tab sprawl you keep losing track of — this is the same idea, purpose-built for the "which one needs me" question. The tradeoff is licensing, not capability: it's dual-licensed AGPL-3.0-or-commercial, fine for solo use, but the copyleft (or a paid license) is something an org has to clear before it goes in the company toolbox. Created March 2026, ~8.7k stars — established and trending, not new this week.

Where herdr watches many agents do many things, this aims many agents at one thing, deep. It's a Go TUI (Apache-2.0, from DoorDash's OSS org) that turns a feature request into a reviewed pull request by running specialized agents across isolated git worktrees through staged phases — knowledge-base build, research, design, roadmap, implementation, review — each a validation gate you can pause at. Install via Homebrew, a prebuilt binary, or source (Go 1.25+); it needs git, the gh CLI, and at least one of Claude Code, Codex, or OpenCode wired up. The live TUI tracks each phase and offers publish/tweak/refactor actions when it lands.

Reach for it on a non-trivial feature where "ask the model to edit files" produces plausible-looking-but-wrong code, and you want the design rigor and review built into the pipeline instead of bolted on after. Delete the ad-hoc "plan, then implement, then ask it to review itself" prompt chain you re-run by hand every time. The tradeoff is that it's young and heavy: 39 stars, and the deeper pipeline profiles (it ships Medium/Large/Moonshot) spend a lot of tokens buying that rigor — this is process-first by design, not a quick fix.

An MCP server that lets your agents search each other's past sessions. It indexes the transcripts each tool already writes locally — Claude Code's ~/.claude/projects, Codex's ~/.codex/sessions — plus your CLAUDE.md, AGENTS.md, and memory/*.md, and exposes recall, search_sessions, and search_memory over them with BM25 ranking, fully offline, nothing uploaded. A tool-specific adapter layer normalizes the different transcript formats into one shape. Install is a single claude mcp add reference -- uvx --from git+... line (Codex, Cursor, and VS Code each get one config snippet).

Reach for it when Codex re-derives a decision Claude Code already worked out an hour ago, because neither can see the other's history — this is the shared-memory layer for a fleet that's otherwise blind across tools. Delete the copy-paste ritual of pasting one agent's conclusion into another's context window. The honest tradeoff: it's one day old (created June 29, 2 stars). The idea is the right one and the mechanism is concretely sound, but this is the bet of the slate, not the proven pick — run it to see where cross-agent memory is heading, don't pin a workflow on it yet.

An open-source CLI plus a local "Studio" web UI that hands an agent production context before it edits a live app: version drift since the last working release, which external providers (APIs, billing, auth) a change touches, and which fixes are repo-code versus dashboard actions that need a human. It ships as skills, so the agent asks "what changed?" first — npx -y viberaven, or npx -y skills add ohad6k/VibeRaven --skill what-broke — then fixes only the surface the evidence supports and flags the rest. Runs with ask/approve/full-access modes for Claude, Cursor, and Gemini agents; on npm as viberaven@1.2.4.

Reach for it when you let an agent fix a bug in a deployed app and it edits blind, with no idea what shipped or which provider it's about to break. Delete the false confidence of a model patching production off a mental model that's three releases stale. The tradeoff: 14 stars, and the "production evidence" is only as good as what you wire in — it structures the judgment about what's safe to touch, it doesn't supply the facts for you.

Off the thread but worth knowing: OmniRoute (github.com/diegosouzapw/OmniRoute, MIT, ~8.1k stars — the pool's biggest product) is a local proxy that routes your coding tools across 50+ providers' documented free tiers using API keys and accounts you authenticate yourself, then compresses inputs 15–95% (command-output filtering plus rule-based prose trimming) before they hit the model. It's legitimate aggregation, not gray-market resale, and it runs entirely on your machine. Not fleet-ops, which is why it sits outside today's thread — but if running many agents at once is about to multiply your token bill, it's the cost valve worth knowing about.