What's new¶
One page per version with what changes for you: new features with their example, fixes that change a behavior and the steps to upgrade. The full detail, change by change, is in each component's Changelog.
-
0.17.0 · in development
The fleet recipe:
deploy_agent()saves it in Parameter Store andAgentFleet.connect(name)reconnects with no project, with the same VMs, warm slots with their gateways and the real scaling instatus. Plus--scale-in-after, opt-in Bedrock key renewal,rayito setup --caps-imagepublishesrayito-base-caps, and fixes to the host logs,--create-secretand OpenCode. No API-breaking changes. -
0.15.0 · 2026-10-21 · current
New
rayito setup: prepares the account and region in one idempotent command (bucket, IAM stack and base image). And deploy the agent you already have to Amazon Bedrock AgentCore Runtime with--target agentcore(rayito agent deploy,deploy_agent()anddeployAgent()): no local Docker, a least-privilege role and the sameinvokeAPI as a MicroVM fleet. AgentCore has no Rayito secrets gateway. With AWS CDK, theRayitoAgentandRayitoMicrovmImageconstructs declare the agent and the MicroVM image in your own app. No API-breaking changes. -
0.14.0 · 2026-10-17
Deploy the agent you already have with zero config, in Python and TypeScript (
deploy_agent(),deployAgent()andrayito agent deploy): adapters for LangGraph, LangChain, deepagents, Strands, Hermes, OpenCode, the Vercel AI SDK or your own app, and typed options for memory, sessions, pool, queue, environment and network. No API-breaking changes. -
0.13.0 · 2026-10-16
The
rayitoCLI, therayito-mcpserver and the local dashboard in TypeScript too (npx rayito), with the same surface, output and exit codes as in Python.rayito --version, UTC timestamps in the Python CLI and a stricter proxy. No API-breaking changes. -
0.12.0 · 2026-10-15
The management dashboard:
Quotas,UsageCostandFleetin the SDK,rayito dashboardlocally and thedashboardstack in your account, with the CloudWatch dashboard, multiple accounts and quota alerts as optional extras. Everything off by default and no API-breaking changes. -
0.11.0 · 2026-10-13
AgentTemplateimages start sooner (create()in about 6 s): therun_codekernel is no longer warmed up by default, a behavior change with a way back (kernel_warmup=True). Per-run artifact keys when publishing images, and the TypeScript pool no longer keeps Node from exiting. No API-breaking changes. -
0.10.0 · 2026-10-08
The agent can use OpenAI, Gemini, Azure OpenAI, OpenRouter, Groq, Mistral, DeepSeek, xAI or your LiteLLM proxy; the timeout and
abort()stop everything it launched; and deepagents works outside Bedrock and counts its tokens. No breaking changes. -
0.9.1 · 2026-10-07
Sandbox.create()with atimeoutof 300 s or less and noidleworks again: the default auto-suspend adapts to the deadline. No API changes. -
0.9.0 · 2026-10-07
Cleanup and fixes: the agent's option D and the gateway exceptions that nothing raised are removed,
connect()recoverssbx.gatewaysfrom another process and the Python SDK honorsAWS_REGION. Breaking changes, with their migration. -
0.8.0 · 2026-10-07
An AI agent inside the sandbox (
sbx.agentwith OpenCode or deepagents), fast start withAgentTemplateand pool warm-up, gateway presets for models and the pricing page. One breaking change inregister_webhook. -
0.7.1 · 2026-10-06
raydcarries its third-party license notices in the release, inrayito-image.zipand in the image (/usr/share/doc/rayd/), andSHA256SUMSis signed. No API changes. -
0.7.0 · 2026-10-05
EFS volumes in your VPC, custom domain (experimental), local testing with Docker and Floci, and security hardening of
rayd, the SDKs, the infrastructure and the release. -
0.6.1 · 2026-10-04
Redeploying an optional stack keeps its parameters,
reincarnate()reapplies all the 0.6 options andraydreaps zombie processes as PID 1. -
0.6.0 · 2026-10-03
S3 mounts, size catalog, events and webhooks, OTLP export, declarative templates, secret gateway and
rayito stack.
Which version you have¶
The Python and TypeScript SDKs and the rayd agent share the MAJOR.MINOR
series. A feature that lives inside the sandbox also needs an image published
with a rayd of that series:
SDK ↔ rayd ↔ image compatibility.
How to upgrade¶
Then, republish your image on the new rayd if you want to use what runs
inside the sandbox (for example volumes=, mounts=, events= or
gateways=, the rayd 0.7.0 hardening, the rayd 0.7.1 license notices in
/usr/share/doc/rayd/ or the rayd 0.10.0 agent timeout that stops its whole
process tree):
Images. Since 0.11, rayito doctor
requires the rayd from the rayd-v0.11.0 tag, and the AI agent
(sbx.agent) also needs an image with its runtime, built with
AgentTemplate (rebuild it
on 0.11.0 so it starts sooner; if you use the scientific stack in run_code,
with kernel_warmup=True). Without republishing, the new SDK keeps working
with your current image, and each feature that needs the new rayd fails
closed with UnimplementedError and terminates the sandbox it just launched.
Going from 0.10.x to 0.11.0 there are no API-breaking changes, but there is a
behavior change in AgentTemplate;
from 0.9.x to 0.10.0 there are no breaking changes
(How to upgrade from 0.9.x); from
0.8.x to 0.9.0, review the
breaking changes (the agent's option D,
GatewayException/GatewayError and, in Python, AWS_REGION); from 0.7.x to
0.8.0, the
breaking change in register_webhook;
from 0.6.x to 0.7.0, the
behavior changes.
Available as experimental¶
Published features, off by default, whose API may change in a minor.
| Feature | Since | How it is used | Why it is experimental |
|---|---|---|---|
| EFS volumes | 0.7.0 | volumes= / volumes, VolumeStore, EfsVolumes (rayito stack deploy efs-volumes), on the optional rayito-base-caps-efs image (rayito image publish --with-efs) |
accepted on real AWS, but they depend on an optional image and on your VPC |
| Custom domain | 0.7.0 | CustomDomain (rayito domain deploy) and register() for each route |
not verified end to end on real AWS (creating the distribution and serving traffic); domain= in Sandbox.create() not wired up |
If you try one, tell us how it went in a GitHub issue.
In development¶
Parts with a pending change that are not available yet. Their option
already exists in create()'s signature so the API does not change when they
arrive, and today it raises UnimplementedError before calling AWS:
| Feature | Option | Status |
|---|---|---|
| Custom domain integrated into the sandbox | domain= / domain in Sandbox.create(), and get_host()/expose() returning the URL |
no committed version; in the meantime, CustomDomain (experimental) |
In the meantime, to expose a port during development,
rayito sandbox proxy.
Designed, no code yet¶
Parts with an already accepted design but not a single line merged into
main: they do not exist in any published SDK, neither as an option nor as
UnimplementedError.
None for now: AgentTemplate, PoolConfig.warmup and the deepagents runtime
were published in 0.8.0.
Earlier versions¶
The notes for releases 0.1.0 to 0.5.x are in the Changelog and in the GitHub Releases.