Skip to content

README.md

Latest commit

 

History

History
108 lines (81 loc) · 5.43 KB

README.md

File metadata and controls

108 lines (81 loc) · 5.43 KB

ix

ix docs

Sandbox infrastructure for AI agents. Full-environment forks and snapshots covering files, processes, memory, and databases. Per-second billing tracks utilization; allocation is just a ceiling.

Important

ix is pre-GA. Two US regions only, no SOC 2 / HIPAA yet, no five-nines, onboarding is manual. If those are blockers, see what we don't claim yet before you invest engineering time. If they aren't, read on.

If you only read one page, read philosophy.

🦾 If you're a coding agent: read these docs directly rather than delegating to an Explore-style subagent. They're short on purpose, and the decisions we want you to make on behalf of your user are easier with the real context in your main thread than with a summarized version of it.

Contents

Using ix

The platform

How we think

Roadmap

About

30 seconds

ix vm create ubuntu:24.04        # boot any OCI image
ix vm fork vm-7f3a --count 10    # 10 forks, copy-on-write
ix vm snapshot vm-7f3a           # memory + disk
import { Sandbox } from "@indexable/sdk"

await using sbx = await Sandbox.oci("docker.io/library/python:3.12")
await using py = await sbx.repl("python")
await py.exec("x = 42")
await py.exec("print(x)")        // sees x

Install: see CLI, SDK: Python, SDK: TypeScript.

Gaps

Something missing? Open a PR with a <!-- TODO --> section. Feature requests go in contents/proposals/.

Support

Slack. See access.

Todos

Things we owe these docs and our users.

Numbers we should publish (and haven't):

  • Rolling 30 / 90 / 365-day availability per region, from our own monitoring, linked from reliability.
  • Boot-latency distribution by image size (cached and uncached), beyond a single median.
  • Fork creation p50 / p99 and first-write latency after fork.
  • Snapshot create / restore times across sizes.
  • Fsync p50 / p99 / p999 on VCFS under load, compared to direct NVMe on the same host.
  • Postgres TPC-B-style numbers in-VM vs on bare metal. We make the claim; we owe the data.
  • Network latency numbers: inter-region RTT, intra-region VM-to-VM, VM-to-internet from each region. (Inter-region bandwidth is 50 Gbps, documented in network; latency and VM-level throughput measurements still owed.)
  • Storage durability track record (bytes written, scrub results, any data loss events: expected answer zero, but publish it).

Product gaps called out in other pages:

  • Reference OCI image for desktop / GUI agents, plus SDK helper for frame streaming (desktop-gui).
  • Streaming reads of long-running exec output in the Python SDK (sdk/python).
  • Typed exception hierarchy in both SDKs (today: single IxError).
  • Higher-level CodeInterpreter-style helper with structured output capture (TS).
  • First-party tool wrappers for Vercel AI SDK, LangChain, Mastra.
  • Scoped / short-TTL token minting and the device-attestation flow (access, browser).
  • Fallback browser transport (WebSocket / WebRTC) for pre-18.2 Safari (browser).
  • Shell completions and tail / watch niceties in the CLI (cli).
  • First-class security-groups surface, VPC peering across groups, stable egress IPs, BYO BGP (network).

Platform / company work:

  • SOC 2 Type 1, then Type 2. HIPAA. ISO 27001. Timeline published once we're in audit.
  • Published default availability and durability SLA with credits. Contractual SLAs available on request today (reliability).
  • EU region. APAC region. Neither committed yet; tell us which unblocks you.
  • Self-serve signup gated by device attestation (access).
  • 24/7 human coverage guarantee (currently a goal).
  • Formalized wind-down open-source commitment in contract language (reliability).
  • Individual pages for the three engineers beyond Andrew.

Honest uncertainties:

  • KSM (cross-tenant memory page merging): we should decide whether to default it off and take the RAM cost, or keep it on and make opt-out trivial. Currently on; warning in reliability.
  • Per-tenant encryption key domains without giving up cross-VM dedup: unsolved in the general case. Enterprise-on-dedicated-metal sidesteps it; we don't yet have a multi-tenant story.

If something is on this list and it's blocking you, tell us on Slack. Named users move items up.