The operating system
for AI agents.
Invisible coordination through the filesystem, built from Unix primitives. Agents don't know it's there. You see everything.
ostk is a local-first operating system for coordinating multiple AI coding agents on a shared codebase. It intercepts file writes at the tool layer, resolves conflicts through optimistic concurrency, and enforces trust boundaries via GPG-based capability pins — all without requiring agents to learn new APIs or coordination protocols.
Agents collide on shared files
Multiple agents editing the same codebase means race conditions, conflicts, and lost writes. ostk intercepts str_replace as a compare-and-swap. No locking, no message queues. Conflicts resolve after the fact.
Agent state scatters on crash
After a crash, you lose context, audit trail, and recovery checkpoints. ostk maintains a durable kernel (.ostk/) tracking agent lifecycle, file generations, and signed checkpoints. Every state transition is logged. Boot from verified state.
Running agents in production means risk
Which Agentfile was executed? Did the code change after signing? ostk enforces trust layers: GPG-sign Agentfiles, bind to HUMANFILE identity, record revocations by content hash. Attestation chains prove what ran.
Classical computer memory mapped to LLM inference primitives. Adding agents decreases per-agent cost.