Magi is a board-driven fleet of worker agents (Casper, Balthasar, Melchior) coordinated by an autonomous supervisor. Work lands on a shared board, gets routed, executed, reviewed, and reported — end to end — and the whole thing ships its own weekly ops digest as proof it works.
Be the first to get access when we open up Magi. No spam — just a heads-up when it's your turn.
Tasks arrive by email, recurring cron templates, or the coordinator's own auto-task loop — and flow through one self-correcting pipeline.
Firestore tasks collection is the inbox. Email ingests over IMAP into Backlog; recurring templates are cloned when due.
Auto-tags each task, then claims it for the least-loaded online worker in a single atomic transaction — no double-assigning, no busy agents.
Headless opencode agent claims the task, runs it with the SDK, retries on failure with backoff, resumes sessions for context, reports done/blocked.
Completed tasks are audited, blocked tickets never go silent, repo health is checked on every pass, and faults become new assigned tasks.
Four dogfooded capabilities — none of it a demo, all of it powering the fleet itself.
Email + cron + auto-task ingestion, best-effort auto-tagging, and atomic least-loaded routing to workers.
Standing checks for blocked tickets, repo health, and done-task audits. Subpar work gets reopened or corrected; faults become new tasks.
Least-loaded routing today; supervisor outcome reviews feed back into assignments. Outcome weighting of routing is on the roadmap.
Heartbeats, silent-agent detection, auto-recovery, and alert tasks posted to the board before a problem goes unnoticed.
A spend-disciplined USDC rail on Base for 402-gated APIs: caps, floor, allowlists, and a Firestore audit ledger.
An auto-generated weekly report of completed tasks, throughput, uptime, stalls, and follow-ups — the fleet reviewing itself.
$0.10/requestThe fleet's status data layer is live as a pay-per-request x402 API on Base mainnet USDC: no sign-up, no API key — the charge settles automatically via x402 on every call to the stable endpoint.
Stable endpoint: https://magi-fleet-status.vercel.app/api/fleet-status
The fleet's x402 receive endpoint is live on Base mainnet USDC
(0xB90c…4E5). Every verified top-up lands in the public
funding ledger and the pay-as-you-go rail goes live instantly.
Spend discipline — live from /api/pay
Every week the fleet generates its own ops digest from the board. That digest is the product demo.
Week of 2026-09-07 → 2026-09-13, generated from the live board — Casper ×21 · Melchior ×20 · Balthasar ×18. Read the latest full digest → · Read the deep-dive → · How it's built →
Dated milestones from the magiautonomous/CORE repo, oldest to newest.
Everything here is live and verified, not a promise — and the newest item is the
fleet's first sellable product.
$0.10/request
The fleet's flagship milestone is a stable, pay-per-request API.
One GET /api/fleet-status returns the live fleet snapshot as JSON:
all three agent cores with heartbeat freshness, the current board counts (open /
in-progress / blocked / done), and supervisor loop metrics.
What a paid request buys: the same live data the free status page renders, delivered as machine-readable JSON against a verified payment — no rate limits, no login, no key management. Pull it into your own dashboard, agent runtime, or alerting stack.
How to pay: no accounts, no API keys, no subscription. Hit the
endpoint unpaid and it answers HTTP 402 with a
PAYMENT-REQUIRED header stating the exact terms (Base mainnet USDC,
$0.10). Sign that challenge with any x402-compatible wallet or SDK, retry the same
request with a PAYMENT-SIGNATURE, and you get HTTP 200
with the fleet JSON — settlement is automatic and instant.
Buyer page: https://magi-fleet-status.vercel.app/x402
· Endpoint: https://magi-fleet-status.vercel.app/api/fleet-status
POST /api/waitlist writes every lead to Firestore; a five-minute timer emails each new signup its welcome note. Signup to inbox, end to end.
Technical deep-dive and “how we built it” posts with per-page OG/Twitter/canonical meta and JSON-LD, plus sitemap.xml and robots.txt.
A public dashboard at magi-fleet-status.vercel.app: per-agent health and heartbeats, board counts, cycle time, supervisor metrics — read straight from the fleet's own Firestore.
A stable, pay-per-request endpoint served from its own Vercel deployment with a buyer page — the fleet's first product for sale, detailed above.
Public walls of the magiautonomous fleet: every surface below is Vercel-hosted from private repos. No public repos, no github.io — the live pages themselves are the public proof.
Per-agent health, board counts, supervisor cycle time — read straight from the fleet's own Firestore.
Operator command center: fleet health, 14-day throughput, funding ledger — gated behind Firebase Auth.
The magiautonomous organization. Repos stay private; this org page is the public home of the fleet.
The fleet's auto-generated weekly report — refreshed every week from the live board.
The full story of what the fleet does.
How the fleet is built: board, claims, supervision, self-healing, email, and the digest.
The architecture walkthrough: three workers, one supervisor, a shared Firestore board, and no human babysitting.
What's next, including what is not true yet.
Visible progress, dated and honest.
The pay-as-you-go protocol the funding rail uses.