Problem

AI agents, training jobs, and evaluation pipelines increasingly operate in parallel over large shared datasets, but conventional storage forces teams into an awkward choice: copy data for isolation, incurring time and storage costs, or let workers share mutable state and risk overwrites, corruption, and poor reproducibility. The killer use case is an AI team running many experiments or agent sandboxes from the same terabyte-scale dataset: each worker needs an isolated workspace, fast startup, durable state, and a clear record of which run produced each result.

Product / Service

Amulet is a high-performance, agent-native Linux filesystem delivered as a managed service and currently offered in private beta. Teams mount a repository like an ordinary filesystem, then create instant copy-on-write branches, snapshots, and versioned commits. A multi-terabyte fork reportedly takes about 3 milliseconds and copies zero bytes initially; mounts complete in under 100 milliseconds, while only changed data consumes additional storage. This lets agents use familiar tools such as Git, grep, editors, and build systems while preserving work after disposable compute disappears and enabling conflict-safe merges and rollback.

The commercial model is usage-based: accounts receive 10 GB free, then pay $0.20 per GiB-month of actively used data, with no separate charges for API calls, transfers, metadata, or forks. That pricing is designed to align costs with the data workloads actually touch rather than every byte retained, while copy-on-write and deduplication keep large numbers of experimental branches close to the cost of the underlying dataset.

Market

Amulet competes in emerging agent-native storage and high-performance filesystem infrastructure for parallel AI workloads. Its alternatives include general-purpose and AI-oriented storage products such as AWS EFS, S3-backed or s3fs setups, JuiceFS, Archil, Modal Volumes, E2B storage, and Daytona storage. The differentiation is not simply faster file access; it is the combination of filesystem semantics, instant isolated branching, provenance, durable state, and conflict-safe publishing for automated parallel writers.

The company is an early-stage YC Summer 2026 startup with a two-person team and is still in private beta, asking prospective users to join a waitlist and seeking access in small batches. Public materials do not disclose customers, revenue, or usage metrics, so the strongest evidence of traction is its YC backing and active product development rather than demonstrated commercial scale; it should be viewed as pre-scale, with revenue status not publicly established.