Amulet vs Archil
Both products give AI workloads a fast, mountable filesystem backed by object storage, at the same $0.20/GiB-month price for actively used data. The difference is what happens after the mount: Archil accelerates access to a bucket, while Amulet versions the work itself — instant forks, commit history, conflict-safe publishing, and per-commit provenance. If parallel agents or experiments write to shared data, Amulet is built for exactly that; if you only need to read an existing bucket faster, Archil is a fine tool.
Amulet
Amulet is a high-performance file system for parallel AI workloads: workspaces mount in under 100 ms, fork instantly without copying data, and keep a versioned, attributable history of every change.
Archil
Archil is an S3-backed filesystem that mounts object storage buckets as a POSIX-like volume with a managed caching layer for fast reads.
Capability by capability
Capability
Amulet
Archil
Mount latency
Amulet
Yes
Under 100 ms, independent of workspace size — no scan or copy on mount.
Archil
Yes
Fast mounts of existing S3 buckets with lazy loading.
Instant copy-on-write forks
Amulet
Yes
Fork a workspace in milliseconds; zero bytes copied until writes diverge.
Archil
No
One shared filesystem view per bucket; no fork or branch primitive.
Version history
Amulet
Yes
Content-addressed commits capture the full workspace state; history is first-class.
Archil
No
Files reflect current bucket state; no commit or snapshot semantics.
Conflict-safe publishing
Amulet
Yes
Publishes are compare-and-swap against an expected base; a stale writer gets an explicit conflict instead of silently overwriting.
Archil
No
Concurrent writers resolve last-writer-wins, as in object storage.
Provenance
Amulet
Yes
Every commit records the actor, session, and run that produced it.
Archil
No
No per-change attribution beyond storage-level access logs.
Restore any revision
Amulet
Yes
Roll a workspace back to any published commit, byte for byte.
Archil
Partial
Depends on S3 bucket versioning configured underneath, per object rather than per workspace.
POSIX filesystem mount
Amulet
Yes
Linux FUSE mount covering standard file workflows — Git, grep, jq, build tools. Live same-file multi-writer access is excluded by design.
Archil
Yes
POSIX-like mount with shared write access delegated between clients.
Read performance
Amulet
Yes
p99 reads of 1.9 ms and 8.4 GB/s sustained throughput per mount.
Archil
Yes
Sub-millisecond cached reads from its managed caching layer.
Credential model
Amulet
Yes
Agents receive short-lived, workspace-scoped grants; object-store credentials never enter the agent process.
Archil
Partial
Mounts authenticate to your bucket; access is scoped at the bucket/IAM level.
Deduplication
Amulet
Yes
Identical file bytes are stored once per account; forks share unchanged data.
Archil
No
Stores what the bucket stores.
Pricing
Amulet
Yes
$0.20/GiB-month for actively used data; API calls, transfers, and metadata free; 10 GB free tier.
Archil
Yes
$0.20/GiB-month for the active cache; API calls, transfers, and metadata free; 10 GB free tier.
Bring your own bucket
Amulet
Partial
Amulet manages the durable backing store; S3-compatible backends are supported.
Archil
Yes
Designed to mount your existing S3, GCS, R2, or other S3-compatible buckets in place.
When to choose Archil
- You want to accelerate access to an existing S3 bucket in place, without changing where the data lives.
- Your workload is read-heavy against data owned by another system, and you don't need history or attribution for writes.
- You want serverless execution colocated with the mount (Archil's disk exec runtime).
When to choose Amulet
- Multiple agents or experiments write to the same data and you cannot afford silent last-writer-wins overwrites.
- You need instant, zero-copy forks — per-agent, per-experiment, or per-eval isolated workspaces from a common base.
- You need to answer "which run produced this file, and from what state?" — per-commit provenance and restorable history.
- You want agents to work with files without ever holding S3 credentials.
Frequently asked questions
Do Amulet and Archil cost the same?
The headline storage price is the same: $0.20 per GiB-month of actively used data, with API calls, transfers, and metadata free, and a 10 GB free tier on both. The difference is what that price buys — Amulet includes versioned history, instant forks, and conflict-safe publishing at no additional cost, and forks cost nothing until their data diverges because unchanged bytes are shared.
Is Archil faster than Amulet?
Both are built for high-performance access backed by object storage. Archil advertises sub-millisecond cached reads; Amulet delivers p99 reads of 1.9 ms with 8.4 GB/s sustained throughput and mounts in under 100 ms. For agent workloads the bigger performance difference is structural: Amulet forks a full workspace in milliseconds without copying data, an operation Archil doesn't have.
Can Archil version my agents' work?
No. Archil presents the current state of a bucket as a filesystem; it does not have commits, branches, or workspace-level restore. Amulet records content-addressed commits with provenance, so any published revision can be inspected, attributed to the run that produced it, and restored exactly.
What happens when two agents write the same file?
On Archil, as on raw object storage, the last writer wins. On Amulet, each agent works in an isolated session and publishes against the base revision it started from — if that base has moved, the publish fails with an explicit conflict rather than silently destroying the other agent's work.