AlfredYour agent says "done." Alfred proves it.

An autonomous coding CLI where "done" means the test suite actually passed — and every hands-off run leaves a signed, tamper-evident receipt you can replay.

Get started

Try it offline — no API key

v0.8.0955 tests3 runtime depsMIT

Verifiable autonomy

A feature_list state machine driven by an objective verify gate — only a captured exit 0 (plus a rubric pass) marks a feature done — with an HMAC hash-chained ledger you can audit and a journal you can replay.

Best-of-breed, not a clone

Memory, orchestration, code intelligence, security, and model routing synthesized from across the ecosystem. Every choice is adopt / adapt / reject, grounded in an ADR.

Security no mainstream harness ships

Lethal-trifecta defense at the content layer — taint fencing, a default-deny egress allow-list, secret redaction, and a dual-LLM quarantine for untrusted input.

Local-first & inspectable

Memory, skills, hooks, journals, and signed ledgers are plain files under .alfred/ that you can cat, grep, and git diff. No cloud control plane.

Watch a verified run

The terminal below replays a real recorded run from examples/demo — implement → objective verify gate → signed ledger row → a one-byte tamper, caught.

examples/demo — replaying a signed run

❯

Pre-recorded replay of a real run — the same record/replay mechanism alfred eval uses in CI. Reproduce it yourself with bun run demo: no API key; the engine, tools, verify gate and signed ledger all run for real.

Where this pays off: overnight autonomous builds · CI gates for agent changes · hands-off runs on untrusted input · receipts you can hand to a reviewer