Routing Claude Code's full stack through Jev for millisecond decisions

Chen @ch3nweiiiOriginally posted 2026.10.16

Open original on X

It is becoming f*cking dangerous: Anthropic documented every piece of this separately, and nobody noticed what they add up to 1. Create a folder on your PC - this is the only world your agent will ever see 2. Launch Claude Code inside it and drop a CLAUDE.md in the root with your rules - it loads into every single session, you never paste context again 3. Create a MEMORY folder and add one rule to CLAUDE.md: write every durable fact there, read it before planning anything 4. Create .claude/agents/ and define three read-only Haiku 5.5 subagents - they explore the files, your main model never burns context on the search 5. Give it a real browser: claude mcp add playwright -- npx -y @playwright/mcp@latest 6. Connect Gmail, Calendar and Drive at http://claude.ai/customize/connectors - they load into the CLI automatically on the same account 7. Wrap the whole chain in .claude/skills/<name>/SKILL.md so the entire thing runs as one command: /<name> the whole setup takes around 20 minutes now the part nobody wires up. every one of those steps has the same slow spot: a tiny fork that doesn't need a thinker. which file. which tool. retry or stop. keep this chunk or drop it. that's where Jev fits. it can't write a word. it returns the choice with a probability in 70 to 500 milliseconds, at $0.042 per million input tokens, output free. that's thousands of decisions for the cost of one LLM call. > let Claude think. > let Jev decide what deserves Claude. > let code do the actual work. what you end up with: the full Anthropic stack running on your own machine - persistent memory, a swarm that explores in parallel, a real browser, your mail and drive, and Jev making the thousands of tiny calls in between the step-by-step setup is in the article below. save this before your next session.

This post quotes another post: x.com/hanakoxbt/status/2110067711363457244