your Max plan runs Hermes

from the episode: X just merged Grok and Cursor into one plan - but there's a catch | LAB 0020

Lenny's AI Builders - LAB 0020

Opus in Slack, on the Claude plan you already pay for.

my Slack bot is called Fwed. it runs on Hermes, the agent framework from Nous Research.

Fwed now runs Claude Opus 5.5 on my Claude Max plan. no API key, and no extra bill.

the trick is one plugin. it makes Hermes run the real Claude app under the hood, so the bill is the plan you're logged in to.

why it matters

back in April, Anthropic blocked Claude plans from running tools like OpenClaw and Hermes. people were angry, and i was too.

then Anthropic paused that change. its help page now says the Claude Agent SDK, `claude -p` and third-party apps still draw from your plan limits.

so if you already pay for Max, your agent can run on it. you don't need a second bill for the same model.

what you get

your own agent on Opus, paid for by a plan you already have.

Opus in Slack for $0 extra. the bot uses your Max plan, not an API key.
it follows your account switch. switch Claude accounts on your laptop, and the bot switches too.
a hard stop at your plan limit. turn extra usage off, and the bot pauses at your limit. it never charges you more.

the real example: Fwed in my Slack

i set this up on 7 October 2026 and checked it again on 13 October, on my own Max accounts.

i asked Fwed in Slack what model it was and how it was running.

it said it was Claude Opus 5.5, running through Anthropic's Claude Agent SDK with the Claude subscription plugin.

LAB 0020 frame at 11:52: Fwed in Slack answers that it is Claude Opus 5.5, running through Anthropic's Claude Agent SDK — source: LAB 0020 episode · 11:52 - Fwed's answer

try it

log in. install Claude Code on the bot's machine and log in with your Max plan: `claude auth login`.
add the plugin. install Nous's Claude subscription plugin in Hermes: `hermes-plugin-claude-subscription-directsdk`.
pick Opus. set the bot's model to `claude-opus-5-5`, with no fallback.
kill extra usage. turn usage credits off in your Claude settings, so a limit pauses the bot.
ask the bot the question below and read its answer.

LAB 0020 frame at 10:25: the route from Slack to Hermes to the Claude subscription plugin to the real Claude app to your Max plan, with the four steps — source: LAB 0020 episode · 10:25 - the route and the four steps

copy this prompt

send this to your bot after the setup.

Lenny’s AI Builders / LAB 0020

what model are you, and how are you running?

you'll know it worked when

the bot says it's Claude Opus 5.5, and that it runs through the Claude Agent SDK, not an API key.

switch Claude accounts on your laptop and ask again. the bot should answer from the new account.

a bad result: it names another model, mentions an API key, or keeps working after you hit your plan limit.

limits and dates

this is one setup, on my own Max accounts, set up on 7 October 2026 and checked again on 13 October 2026.

Anthropic's help page says the Agent SDK can run on your plan limits. it also says an account is for one person. my team uses Fwed - that's my call, and you make your own.

prices and rules change fast. check Anthropic's help page before you set this up.

LAB 0020 went public on YouTube on 14 October 2026 at 10:53 am Perth time.

the rest of this LAB

X merged Grok and Cursor into one plan. the new X Subscription puts Grok, Cursor and X Premium in one plan: Plus is $8 a month, Premium $30, Super $100 and Ultra $200. it's opt-in, and Grok Bot comes with Premium and up. the catch: Claude and ChatGPT inside Cursor now bill at the provider's API list price. i switched, and my plan check says it saves me about A$74.60 a month. my take: check your Cursor models first.
Meta's Muse Wonder lands tied 2nd. on the Artificial Analysis Intelligence Index, Meta went from 48 with Muse Spark to 52 with Muse Wonder and 56 with Muse Wonder Full, only behind Opus 5.5. it costs $4 per million input tokens and $20 per million output tokens, with a million-token context window. my take: test it this week - it's the first Meta model i'd pay to use.
Hermes and OpenClaw are back on Claude. on 7 October, Max plans got API credits that work in your own code and in third-party tools like OpenClaw and Hermes. Pro gets none, and the credits don't cover Claude Code. and the Agent SDK still runs on your plan limits - that's the practical above.
OpenAI day 7: 500k context. GPT-6.1 Sol's default context in Codex and Work went from 272k to 500k. there are no long-context fees when it draws from your subscription. my take: long chats just got longer, and it compacts less.
Google's notetaker never leaves your Mac. Google AI Edge Foresight for macOS uses local Gemma 4 models to turn voice notes into documents. nothing is sent to a cloud model. my take: it's a Granola that never leaves your Mac.
Nadella: treat models as insider risks. Satya Nadella's essay says to treat every model like an employee with keys. one rule: use different models to check each other. my take: give your model the keys - just don't let it check its own work.

sources

next step

watch LAB 0020 on YouTube - the practical starts at 10:18.

want the next one? get the next field notes. optional - these notes stay free to read.