Hermes Agent Milestones, Claude Code 2.1.142 & Kimi Web Bridge
Welcome back. Let's get into it.
NousResearch hit 150,000 GitHub stars on their Hermes Agent project this week — and the team is not being quiet about it. Cofounder and lead engineer Teknium posted that Hermes is almost double-ing OpenClaw's daily token volume — just three days after first surpassing it. His words: "I don't love to gloat, but…" — and then he gloated. Fair enough when the numbers back you up.
Speaking of Hermes Agent — if you haven't looked closely at how its memory system works, it's genuinely interesting engineering. Akshay Pachaar broke it down in detail. There are three tiers. Tier one is a tiny markdown file called MEMORY.md — capped at 2,200 characters — that gets injected directly into the system prompt alongside a USER.md file. When it hits about 80% capacity, it consolidates automatically. Tier two is full-text session search through SQLite with FTS5 — we're talking roughly ten milliseconds across more than ten thousand documents — and the LLM summarizes the top hits. Tier three is where it gets modular. Eight pluggable memory providers. Things like Honcho, which does dialectic user modeling across twelve identity layers. Holographic, which runs local-first HRR vectors with zero external calls. Supermemory, which handles context fencing. The whole system composes through a five-step cycle with a roughly 300-second reflection nudge built in. That's a lot of memory architecture packed into one agent framework.
Now — while Hermes is surging — OpenClaw is still shipping. Version 2026.5.12 dropped on May 14th. The headline changes: OpenAI setup now defaults to Codex login, there are runtime fallbacks and stalled-stream recovery baked in, Telegram polling survives stalls, and startup is leaner and faster. And separately, a developer named shenhao-stu dropped a repo called openclaw-agents that spins up nine specialized AI agents for OpenClaw in a single command. One command, nine agents. That's the kind of setup script people actually share.
Alright — shifting over to Claude Code — version 2.1.142 landed with 24 changes, and the big one is a new set of flags for the `claude agents` command. We're talking `–add-dir`, `–settings`, `–mcp-config`, `–plugin-dir`, `–permission-mode`, `–model`, `–effort`, and `–dangerously-skip-permissions` — all designed for running background sessions. Grep now defaults to ripgrep. The macOS daemon detects clock jumps after sleep to stop session loss. And fast mode now runs on Opus 4.7 by default.
Also worth noting — there's an official Anthropic plugin called `claude-code-setup` that analyzes your repo and auto-generates hooks, MCP servers, skills, subagents, automations, and workflow configs. You install it with `/plugin install claude-code-setup@claude-plugins-official`. That's a real time-saver for anyone setting up a new project from scratch.
Meanwhile, a Reddit user called CohleM built nanoclaude — a Claude Code clone — from scratch in one hour. The repo is up on GitHub. One hour. Whether that's impressive or slightly alarming probably depends on how much you paid for your Claude subscription.
Which brings us to the debate that exploded this week. Developer hiarun02 asked the obvious question — why pay $100 a month for Claude Code when DeepSeek V4 handles the same coding task for $5? The post lit up with over a thousand likes and more than 300 replies. The counterarguments centered on reasoning quality, workflow reliability — and for a lot of people — data privacy concerns around routing code through Chinese infrastructure. It's a real tension, and it's not going away.
On the tooling side — Moonshot AI's Kimi team launched something called Kimi Web Bridge. It's a Chrome extension that lets Claude Code, Cursor, Codex, and Hermes agents perform human-like browser interactions — search, scroll, click, type — all of it. The demos show it auto-creating full Google Forms, cloning websites using K2.6 multimodal vision, and pulling multi-platform search results into spreadsheets. This is the kind of glue layer that makes coding agents actually useful outside of a terminal.
And finally — the ex-cofounder and CTO of Spark AI, a YC W24 company, Taehoon Kim, has joined Cursor to work on future coding agents. When YC founders start moving into established AI coding companies, that says something about where the gravity in this space is sitting right now.
That's your AI digest for 15 May 2026.