
A durable shared disk for AI agents that sits on top of your object storage, with serverless sandboxes that bill per tool call and a free tier.
Archil gives AI agents a persistent shared disk on top of object storage, so teams stop writing S3 sync pipelines. Founded in 2024 by an ex-Amazon EFS and Netflix storage engineer, it now sells Free, Developer, Team and Enterprise plans, and the paid tiers bill usage against a monthly minimum.
Idle-free serverless sandboxes billed by the minute, running directly against a cached S3 mount, so teams skip both ETL pipelines and always-on compute.
The founder's Amazon EFS and Netflix storage background and a custom caching protocol give some technical edge, but FUSE-based S3 mounting is a well-trodden space (s3fs, goofys, Mountpoint) that a well-resourced competitor could replicate.
Main trade-off
Eventual consistency between Archil and the source S3 bucket, plus AWS-only serverless execution, limits fit for workloads needing instant cross-system consistency or non-AWS serverless compute.
| Plan | Monthly price | What it includes |
|---|---|---|
| Free | Free | 5 file systems of up to 10 GB each, 1,000 serverless tool calls a month, 1 hour of sandbox compute a month, 5 concurrent sandboxes, 30-minute sessions, Hard caps with no overage |
| Developer | $10/mo | $10 monthly minimum, usage counts toward it, 1,000 file systems and 20 concurrent sandboxes, External mounts and bring your own bucket, Network controls and private sandbox images |
| Team | $500/mo | $500 monthly minimum, usage counts toward it, 25,000 file systems and 250 concurrent sandboxes, 3 workspaces, SOC 2 and a shared Slack channel |
| Enterprise | Custom | Custom minimum and unlimited capacity, SSO, SCIM and advanced API key permissions, BYOC or on-premises, PrivateLink or VPC peering, HIPAA with BAA and DPA, availability SLA, 24/7 support |
Free: $0 with hard caps and no credit card; storage, tool calls and sandbox hours all stop at the cap, and external mounts are not included. Developer: $10 monthly minimum, and usage counts toward it. Team: $500 monthly minimum with SOC 2 and a shared Slack channel. Enterprise: custom minimum, adds SSO and SCIM, bring-your-own-cloud, PrivateLink, an availability SLA, HIPAA with BAA and DPA, and 24/7 support. Usage rates on Developer and Team: $0.05 per 1,000 serverless tool calls, $0.20 per GiB-month active storage, $0.025 per GB-month archive storage, $0.055 per vCPU-hour, $0.02 per GiB-hour memory, $0.07 per GB leaving Archil.
| Feature | Free | Team | Developer | Enterprise |
|---|---|---|---|---|
| File systems | 5 (10 GB each) | 25,000 | 1,000 | Custom |
| Concurrent sandboxes | 5 | 250 | 20 | Unlimited |
| External mounts and own bucket | — | ✓ | ✓ | ✓ |
| SOC 2 | — | ✓ | — | ✓ |
| SSO and SCIM | — | — | — | ✓ |
| HIPAA with BAA | — | — | — | ✓ |
Archil is a hosted disk that mounts Amazon S3, Google Cloud Storage and Cloudflare R2 buckets as a POSIX filesystem, with vendor-measured 0.5 ms cache-miss reads. It adds serverless Bash, Python and Node tool calls billed per call, plus persistent sandboxes, for AI agent and ML data work.
Object storage as a POSIX disk
Mounts Amazon S3, Google Cloud Storage, Cloudflare R2 and S3-compatible stores as a real directory, with hot data on SSD and a published 0.5 ms cache-miss read for 4 KiB.
Serverless tool calls
Each disk.exec call runs Bash, Python or Node in a fresh microVM against the disk, with execution included in the call price and files kept for the next call.
Persistent sandboxes
Long-running servers boot from standard OCI images and bill for the CPU and memory they hold while running, with a concurrency cap that rises by plan.
Checkpoints and branching
Lets separate agent runs work on independent branches of one data set; on bring-your-own-bucket disks this is an Enterprise-only feature.
Sandbox egress firewall
Paid plans can deny all egress or allow listed domains and CIDR ranges; Free uses a managed allowlist of registries, Git hosts and model APIs.
Kubernetes CSI driver and Terraform provider
Provisions volumes in Kubernetes clusters and manages disks as infrastructure-as-code, with documented guides for Modal, Vercel Sandbox, Fly.io Sprites and Cloud Run.
Bring your own bucket
Paid plans write each file to your own bucket as a plain object, so removing a disk leaves the data where you can still reach it.
Four generally available regions
Disks run in AWS us-east-1, us-west-2 and eu-west-1 plus GCP us-central1, with other regions on request.
Most users expect these and they are not currently supported.
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| $10/mo | Yes | Web | 3.9 | |
| $150/mo | Yes | Web | 3.9 | |
| $250/mo | Yes | Web | -- | |
| -- | Trial | Web | -- |
Proprietary hosted service; TypeScript SDK (archil-sdk) is open source under MIT
Archil has four plans: Free at $0 with hard caps, Developer with a $10 monthly minimum, Team with a $500 monthly minimum, and custom Enterprise. Metered usage counts toward the minimum, so $6 of Developer usage bills as $10 while $18 bills as $18. Paid usage runs $0.05 per 1,000 tool calls, $0.20 per GiB-month active storage and $0.07 per GB out.
Yes. The Free plan needs no credit card and includes 5 disks of up to 10 GB, 1,000 serverless tool calls and 1 hour of sandbox compute a month. Work stops at the limit instead of billing. Free has one workspace, no external mounts and sessions of up to 30 minutes.
E2B covers short-lived sandboxes without a persistent disk, and Modal covers general serverless compute. Archil is the pick when agents must keep and share files between calls, because the disk stays after each tool call ends. Teams that only need a free FUSE mount of a bucket can look at AWS Mountpoint for S3 or s3fs-fuse.
Not on paid plans that use bring-your-own-bucket. Each file lands in your own bucket at its normal path, and deleting the disk leaves those objects untouched. Managed disks use an internal format, which you export with any S3 client or by mounting the disk and copying files.
Create a free account at console.archil.com, install the TypeScript SDK (the disk package) or the Python SDK, then create a disk and call disk.exec to run Bash, Python or Node against it. For agents, the vendor documents a Vercel AI SDK helper that exposes the disk as tools. A mount takes about 50 ms according to the vendor.
Hunter Leath founded Archil after nine years in cloud storage, first as a founding engineer and senior product manager on Amazon EFS, then leading cloud storage at Netflix. Founded in 2024 as Regatta Storage and renamed Archil in 2025, the company now describes itself as "the serverless runtime for intelligence": durable disks plus on-demand sandboxes for AI agents, evals and data pipelines.
The core idea is a disk that keeps hot data on SSD and cold data in object storage. You can start with an Archil-managed disk or connect an existing bucket on Amazon S3, Google Cloud Storage or Cloudflare R2 (MinIO, DigitalOcean Spaces, Wasabi and Backblaze B2 work through the S3-compatible path). With your own bucket every file is an ordinary object any S3 tool can read without Archil. On top of the disk, a disk.exec call starts a fresh microVM for your shell, Python or Node code, and whatever it writes is there for the next call. Persistent sandboxes cover servers that must stay up, and checkpoints and branching let separate agent runs work on independent copies of the same data.
Pricing was reworked since the last Hokai review. The old Developer and Team plans are gone: there is now a hard-capped Free plan, a Developer plan and a Team plan that each carry a monthly minimum rather than a subscription fee, and custom Enterprise terms. Usage is metered per tool call, per GiB of storage, per sandbox CPU-hour and per GiB-hour of memory, and the exact rates sit in the pricing section of this page. The product plugs into Kubernetes through a CSI driver and into infrastructure-as-code through a Terraform provider, and the vendor lists guides for Modal, Vercel Sandbox, Fly.io Sprites, Cloudflare Containers and Cloud Run.
Where it fits matters. Archil is aimed at ML engineers training on large S3-hosted data sets, teams running fleets of coding or research agents that need shared, persistent storage, and engineers who maintain custom sync pipelines. If you only need short-lived, stateless code execution, a pure sandbox such as E2B is simpler, and if the job is general serverless compute rather than a filesystem layer, Modal is the better fit; teams that mainly want hosted inference for open models are better served by Together AI. Small projects with a local data set gain little from a mounted object-store disk and may find the paid plans' monthly minimum unnecessary.
Archil is reported to have raised $18M in total, led by Felicis Ventures at seed and Standard Capital in an April 2026 Series A, and the vendor lists Clay, Gamma, Gumloop, Depot and microAGI as customers. The speed numbers on its homepage (small cache-miss reads against typically tens of milliseconds from S3, and npm ci for 598 packages in 7.6 s against 86.3 s on Amazon EFS) are vendor-run measurements, and the vendor points to a public script and an EFS comparison page so buyers can reproduce them.
Product facts on this page (pricing, plans, integrations, compliance) are read from Archil's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.