Open source, and built for one person: you. Edward runs on your Mac, knows where he heard what he knows, keeps the promises he makes, and works in a sandboxed computer of his own.
Capabilities
Every memory records who said it, what it is about and how sure he is. Ask how he knows something and he shows you. Only your own words become his standing instructions.
A persistent shell, a file editor and a headless browser, sandboxed on your Mac: he can’t touch his own running code, read his keys or reach your database.
When he says he’ll check back, he schedules it and comes back. Whatever he does on his own gets a journal entry you can mark right or wrong.
Opt-in noticing reads your messages for what matters (a deadline closing, someone waiting on you) and says it once, through one door, outside your quiet hours.
Write @Edward in any chat and he answers there, from your own account. Friends can ask too, inside limits; unknown numbers wait for you.
A daily spend cap in dollars, a cap per turn, a secret scan on everything that leaves, and approvals when you want them.
How it works
In the app, by voice, or with @Edward in iMessage. The same Edward, with the same memory, everywhere.
He recalls what he knows, searches the web, works in his own shell, and shows you every step as he goes.
He answers, files what he learned with where it came from, and comes back when he said he would.
Where he shows up
An app you can install on your phone, live voice, an iPhone widget he fills himself, and push notifications that wait out your quiet hours.
Open Source
You need a Mac with Apple silicon that stays on, Homebrew and an Anthropic API key. Edward keeps what he knows in your own Postgres; the one outside service he needs is the Claude API.
git clone https://github.com/ben4mn/meet-edward.git
cd meet-edward
./setup.sh # asks your name, API key and time zone
bash scripts/install-launchagents.sh
# then, on this Mac, open http://localhost:3000
cat backend/.setup_code # the code for your first passwordUnder the Hood
Edward is a long-running service on your Mac that can read your messages and act as you, so he comes with rails, receipts and limits.
pgvector and keyword search, blended 70/30, over a small embedding model that runs on your Mac. Provenance on every row, and upkeep that never deletes.
The same tools in chat, voice, pings and wakes. What a turn may do is its policy, and a refusal comes back as an honest error, never a hidden tool.
His shell can’t read his keys, write his own code, reach his API or databases, or script your apps. It is not a containment boundary, and the docs say so.
Follow-through on promises, a silent nightly, one small budgeted run of free time a day, and a journal of what he chose to do, instead of a timetable.
Sonnet for chat and wakes, Haiku for memory work, Opus for evaluations. Every call priced and counted against a daily cap.
A cheap noticer reads what arrives, a gate in plain code decides, a verified speaker says it once. Quiet hours are enforced in code.
LaunchAgents, a bottom-up watchdog, verified nightly backups, and staged frontend deploys with a one-command rollback.
CLAUDE.md and a page per subsystem, kept honest by a doc lint, so Claude Code can help you set him up, run him and change him.