Build agents.
Not infrastructure.

Subako is the cloud runtime for the AI agents inside your product. Long-running sessions, your own tools, secure code execution, credentials and skills, run as one service. Use the models Subako provides, or your own.

自社プロダクトの中でAIエージェントを動かすためのクラウド基盤。長時間実行のセッション、自作のツール、コード実行、認証情報、スキルを、ひとつのサービスとして動かします。モデルは Subako のものも、自社の契約のものも使えます。

Get started

Subako, the film. 1:26, with sound.

Why Subako

A model alone stops at an answer.

To do real work, an agent needs a runtime around the model: something that keeps the loop running, holds state across a dropped connection, executes code safely, carries credentials without exposing them, and knows the procedures. Today every product team builds and operates its own.

エージェントが仕事を終えるには、モデルの外側に実行環境が要ります。ループを回し続け、切断をまたいで状態を保ち、コードを安全に実行し、認証情報を漏らさずに運び、手順を知っている部分です。いまは、それをプロダクトごとに各チームが自分で作って運用しています。

A product window wired to a tangle of runtime components
Today. Each team builds, hosts and secures the runtime, for every product.
A product window connected by a single line to the Subako runtime
With Subako. One runtime, hosted and managed, behind one SDK.

What you get

Everything agents need to run.

Six components, one service. Each one is production-grade on its own, and you never operate any of them.

6つのコンポーネントを、ひとつのサービスとして。それぞれが本番品質で、どれも自分で運用する必要はありません。

harness
sessions
sandbox
vault
skills
functions
harnessLong-running executionRuns the reasoning and tool-call loop until the job is done. Context is compacted automatically, and you can steer or stop a run mid-way.
sessionsState and reconnectionAn append-only event log is the only source of truth. Drop the connection and resume from the last event. A crash rebuilds from the log.
sandboxA Linux machineA Linux machine of the session's own, for programs that only run on Linux. It boots on the first command and, after five idle minutes, is suspended with its workspace archived.
vaultCredentialsPer-secret envelope encryption with KMS. Write-only: a secret never appears in a listing and never reaches the model. Per-user vaults let an agent act as that person.
skillsKnowledge and proceduresVersioned bundles of prompts and scripts in the Agent Skills format. Pin a version or follow latest. Metadata is injected into the system prompt for you.
functionsYour own toolsTools you write in Rust, TypeScript, Go or any language that builds to WebAssembly, run by Subako beside the session. No server to run, no network unless you grant it, and credentials the code never sees.

Security by construction

The model never sees a secret.

Prompt-injection defence is structural, not a content filter. Least privilege is declared per tool, approvals gate the risky calls, and credentials stay outside the engine.

プロンプトインジェクション対策はコンテンツフィルタではなく構造で行います。ツール単位で最小権限を宣言し、危険な呼び出しは承認で止め、認証情報はエンジンの外に置きます。

A broker between the agent and your tools

The engine receives tool names, schemas and a bearer issued per run. Real endpoints and credentials are injected by the broker. There is no secret for the model to leak.

Policy per tool, set per grant

Set Allow, Ask first or Deny for each MCP server, sandbox and function, and override it per tool by exact name. Tools you did not name, including ones the server adds later, follow that setting.

Approvals that hold nothing

A call that needs a human parks the run and releases its lease. It resumes when the person answers, yes or no. A call nobody answers within 24 hours is denied, and the agent is told.

Encrypted per secret, audited per action

Every secret is envelope-encrypted with its own key under KMS. Every change lands in an append-only audit log. Access across organisations returns 404.

Any model

Any model. Any provider.

Run on the models Subako provides, paid from credit, or connect your own Anthropic, OpenAI or Amazon Bedrock account and pay the provider directly for the tokens. With your own provider, requests keep the provider's own format, so a new model can be used by its ID the day it ships. Tools speak MCP, skills follow Agent Skills, and identity is OpenID Connect.

Subako が用意するモデルをクレジットで使うことも、自社の Anthropic・OpenAI・Amazon Bedrock の契約につないで、トークン代をプロバイダに直接支払うこともできます。自社のプロバイダにはプロバイダ本来の形式のまま中継するので、新しいモデルも出た日からその ID で使えます。ツールは MCP、スキルは Agent Skills、認証は OpenID Connect と、オープンな標準でつながります。

claudegptbedrockopen-weight
MCPtoolsThe MCP servers you already publish become the agent's tool surface.
OIDCidentitySubako is an OpenID Connect issuer. Your AWS, Anthropic or OpenAI account trusts a workspace, and no key changes hands.
Agent SkillsknowledgeSkills written for other harnesses carry over unchanged.
modelscredit or your ownThe models Subako provides, from credit, or your own provider over the Anthropic or OpenAI Responses API.

Deployment

Cloud or private.

The same Subako runs as a shared cloud, as an environment dedicated to you, or on your own infrastructure. What changes is where it runs, who operates it, and how it reaches its models.

同じ Subako が、共有のクラウドとして、お客様専用の環境として、あるいは自社のインフラで動きます。違うのは、動かす場所と運用する人、モデルへのつなぎ方です。

Cloud

Start right away. Tenant isolation per workspace, stateless replicas, sandboxes that start on demand.

Dedicated

An environment for you alone, in an AWS environment we create for you or in your own AWS account. Models, data residency, network and sign-in are shaped to your requirements.

Talk to us

On-premises

Subako on your own infrastructure, operated by your team with our support.

Talk to us

The infrastructure for running AI agents.