Orvena
problem
Some teams want an agent platform but need conversations, memory, attachments and secrets to stay on their own hardware, behind their own auth and network rules. A multi-tenant cloud service can’t offer that.
solution
Orvena runs as one instance per customer host. A Rust daemon stores actors, queues, alarms and vectors in FeOxDB, and a TypeScript layer runs the agents, tools, sandboxes and privacy pipeline on top.
results
Turns survive the daemon and the API being killed mid-turn, and an automated restart test checks that no action runs twice. 105 Rust tests and about 1,000 TypeScript tests cover the rest.
Architecture
orvenad is a stateless HTTP layer over two FeOxDB files: one for state that can’t be rebuilt, one for vectors. Execution stays in the TypeScript API, which treats actors as storage identities rather than running objects.
storage
State transitions without multi-key transactions
FeOxDB offers compare-and-swap on a single key, not multi-key transactions. So every message has one authority record with a revision, and the ready, leased and dead indexes are projections of it. A transition writes the new projection, swaps the authority, then deletes the old projection. Readers ignore any projection whose revision doesn’t match, so a crash between any two steps leaves a state recovery can resolve.
concurrency
Fencing tokens for actor state
A worker that stalls past its lock lease can wake up and write stale state. Each lock grant takes a new fence from a counter that is flushed before the grant, and every state write carries the fence it was granted. A write with an old fence is rejected, so only the current holder can change the actor.
on the phone
Orvena on iPhone
A separate app built on the same design at phone scale: a persisted turn queue per conversation, a replay outbox and lazily loaded skills around MLX models running on the device. It shares no code with the hosted runtime.