Wren: a home-made Claude Duet

Anthropic's Claude Duet is a background agent that checks in each morning, does long jobs on its own and asks before it sends or pays. It's waitlist-only for now. We had already built most of the same thing at home, out of Claude Code, a personal server and a handful of open tools. We call it Wren. This is what it's made of, how it compares, and what it taught us.

How it grew

• June 2026. Session tooling (crss) lets Claude Code run for days on a server and be reached from a phone.
• August. A shared message bus lets Claude, Codex, ChatGPT and other agents hand each other work.
• October. DispatchBench picks the cheapest model for each job, and the orchestrator gets a name, a persona and a memory: Wren.
• 26 October. A 07:00 Telegram brief, and a phone approve/deny gate, tested live both ways.

How you talk to it

From anywhere. Telegram (typed or voice notes), the Claude app's Remote Control and voice chat, the desktop app, or ChatGPT, which reaches the server through a connector. Every door opens onto the same server and the same memory, so you can run several Wren threads at once.

What it's made of

• Claude Code on a home server, signed in to a Max plan. Opus does the planning and checking.
• crss: starts, messages, hands off and repairs long-running sessions. Every agent on the box runs on it.
• token-reduce and QMD: read only the slice of a repo or notes folder a task needs. QMD is a local search engine over our markdown, re-indexed every six hours.
• DispatchBench and the delegate skills: send each job to the cheapest model that passes. Codex builds, Haiku sweeps and reads, Fable gives a second opinion, and Kimi or Grok can be added through the same skills whenever we subscribe.
• A soul: a short persona file every session loads, plus a cheap reviewer that saves one lesson before each context reset. Wren keeps getting better at its own job.
• Shared memory: a gbrain knowledge base plus plain git-tracked notes that Claude, Codex and the other agents all read and write.
• The bus: a small mailbox database where agents post tasks and replies to each other, including Hermes and OpenClaw.
• Phone bridges: the Hermes gateway and takopi for Telegram; Claude Remote Control for the app.
• ChatGPT's apps, through Codex: about 25 connectors, including Interactive Brokers, Questrade, Alpaca, Binance and CoinGecko, plus GitHub and site hosting.
• Market data on the server: Databento, Yahoo Finance, FRED, SEC EDGAR, Helius for Solana, and web search.
• Daily jobs: about 25 scheduled tasks, the 07:00 brief that summarises them, and the approval gate.

Against Duet

Fourteen features, checked against the Claude Duet product page on 26 October. Our own assessment, not a benchmark: Duet is more polished and works out of the box.

What it costs

Wren runs on the Claude Max plan we already pay for, plus Codex and ChatGPT subscriptions we already had. Duet comes with Pro and Max plans and shares their limits. Here's what the same work would cost at API list prices, so you can judge for yourself.

What we learned

• Expensive head, cheap hands. The strong model plans and checks; cheaper models do the reading and building.
• A "done" is a claim. Wren re-runs the check before it believes a helper, and so should you.
• Borrow openly, then add only what fixes a repeat failure. We borrowed openly from Hermes and OpenClaw: the persona, the lesson saved before each reset, the rules for what not to remember. After that, a new habit gets in only when it fixes a failure we keep hitting.
• Remove scaffolding as models improve. Each new Claude model needs less hand-holding, so hooks and rules get re-tested and dropped once they stop paying.
• Own the box, own the gates. You decide what runs, where, and what needs your yes first.

The whole thing on one page

In short: one strong model on your own server, reachable from your phone or ChatGPT, handing work to cheaper models and other agents, and asking you before anything leaves the building.