Agents produce work faster than human attention can follow it, and the tooling has rushed to fill the gap with more generation. These five launches go the other way: they are about the accounting layer. Where did the sessions go, is the model you picked actually good at your code, are you about to hit a plan limit, and is the UI the agent just shipped any good? Every one of them answers a question the agent itself cannot be trusted to answer.
Sessions that survive the agent
Agent Archive saves your Claude Code, Codex and Cursor sessions to your own S3 or Cloudflare R2 bucket. The handoff command picks a session and carries its conversation context into a different coding agent or a different machine, which matters more than it sounds: you lose your thread the moment you switch agents, and then spend twenty minutes working out where you were. The list command browses history across agents, filterable by topic, project, branch or PR number, and the stats command breaks down session counts, tokens and estimated spend by agent, model and project.
It ships as release binaries for macOS on Intel and Apple Silicon and Linux on x86-64 and arm64, MIT licensed, and the archiving is opt-in — connect your own bucket and transcripts start uploading. The honest limitation is that it is a CLI that lives beside your agent rather than inside it, so the value shows up when you actually switch agents or machines, not on day one.
A calm home for Claude Code
Perch is a free beta Mac app for Apple Silicon that gives your Claude Code work one place to live. It resumes the sessions you pick exactly where they left off, detects linked PRs across repos and worktrees automatically, and — the part that matters most — stays quiet while you are in flow. You hear from Perch only when Claude, a reviewer or CI is waiting on you, and deep focus mutes everything else so you can stay with one piece of work.
Each piece of work keeps a short checklist where every step has an owner: you, Claude, or a teammate. Claude writes a one-line status as it works, and each PR shows who you are waiting on. Privacy is the strongest part of the pitch: contexts, notes and status live in a local database, there is no account and no Perch servers, no analytics, and crash reports are off unless you turn them on. It requires the Claude Code CLI, and GitHub features need your gh login.
Usage limits where your eyes already are
LLM Usage Widget is a small desktop widget with a single job: show your Claude and Codex plan usage limits where your system tray icons live. On Linux it appears in the top bar next to Wi-Fi and sound, on macOS in the menu bar, on Windows on the taskbar and in the tray. The dropdown lists every limit with its percentage and reset countdown, and the bars turn amber at 70 percent and red at 90.
The detail that makes it work is that there is nothing to configure: it reuses the logins that Claude Code and the Codex CLI already store on your machine, reading credentials without modifying them and sending each token only to the provider that issued it. When a plan hits 100 percent it switches to showing what is left of your usage credits instead, and a failed refresh keeps the last numbers marked stale rather than lying to you. MIT licensed across all three desktops, and Claude and Codex only.
Benchmarks built from your own pull requests
SelfBench attacks the problem that public benchmarks answer the wrong question: they tell you how a model does on someone else's code. It turns your repository's merged PRs into private eval tasks — for each PR, the task is rebuilt from the commit before the change, with the PR's own request as the instruction and an authoring agent writing hidden tests and a reference solution. Then agents and models run against them, and accuracy is plotted against cost with the Pareto frontier drawn through the settings nothing else beats on both.
The acceptance gate is the interesting part: a task is accepted only if the tests fail without a solution, pass with the original implementation, pass again on a rerun, and survive an independent review. Every accepted task is a native Harbor task. Repositories released publicly get a live leaderboard on selfbench.dev — next.js, supabase, sentry, posthog and others are already there. Batch generation takes hours and happens in the hosted web app, so this is a commitment rather than a quick check.
A critic for the UI your agent shipped
Pristine is an agent skill that critiques, designs and polishes web UI. It captures the real page — light and dark, desktop and mobile — and scores it out of 10 across 11 dimensions through the eyes of the personas your product actually serves, with the exact fix for every finding. The polish command then keeps the look and raises every detail, and 16 named lenses force a direction: bolder, typeset, distill, clarify, harden and so on. The init command reads your code and writes a DESIGN.md with product, personas, brand and design system.
Eight published before/after pairs show the full critique and every change, which is more transparency than most tools in this space offer. It works with Claude Code, Codex, GitHub Copilot, Cursor and any agent that reads SKILL.md, and it needs a browser tool such as Playwright MCP to capture pages. Apache-2.0. The caveat is worth stating plainly: the score is the agent's judgement, not user research.
The through line
Each of these refuses to take the agent's word for it. The archive records what happened, SelfBench measures which model earns its cost, the widget tells you before you are cut off, Pristine judges the output, and Perch tells you what needs you now. Individually they are small utilities; together they are the difference between running agents and running a team of them.