Controlled environments where AI agents work safely and improve

Granular wraps around your tools and gives agents a programmable environment to run complex logic with state and visibility.

See agents built with Granular
Build your agent

Problems

Why agents fail in production

Real failures that happen when agents lack a proper execution environment.

Runs out of context

Sandboxed agents persist state independently of the prompt. They remember context across sessions without hitting token limits.

Acts out of guardrails

Sandboxed agents require explicit approval before destructive operations. Every risky action is validated before execution.

Loses conversation context

Sandboxed agents maintain persistent memory across conversations. They remember past work and can seamlessly resume from where they left off.

Struggles with complex logic

Sandboxed agents execute code in a real environment. They can build state machines, handle retries, and orchestrate complex workflows without losing progress.

Guesses instead of asking

Sandboxed agents pause and generate interactive forms to collect missing information. They ask before assuming.

Can't explore side effects

Sandboxed agents can query databases, inspect schemas, and test operations before executing. They explore 'what if' scenarios to understand full impact.

Setup

Build a sandbox around your agent's tools



1

Granular wraps around your tools

Granular wraps your existing tools with a sandbox layer.

Your agent and your tools stay the same.

const server = new McpServer({
  name: 'my-server',
  version: '1.0.0'
});

// .... other tools ...

const sandbox = new McpSandbox({
  graph: 'test-sandbox',
  privateKey: process.env.GRANULAR_PRIVATE_KEY,
});

const wrappedServer = await sandbox.wrap(server);

All tool access goes through the sandbox.

2

Add sandbox capabilities

You extend what the agent can do by adding capabilities to the sandbox: memory, durable state, processes and state machines, rules and approvals, UI components, introspection and traces.

🧠

Memory

●Facts memory●Shared knowledge●File system storage●Vector storage●Key value storage
3

Define permissions and approvals

You control which actions are allowed, restricted, or require approval.

4

The sandbox becomes the agent's environment

The agent becomes a coding agent, writing code that programs against the sandbox environment. The sandbox executes that code and drives the tools.

Capabilities

How agents work inside the sandbox

Inside the sandbox, the agent writes code against the sandbox environment.

That code handles logic, state, rules, and UI inside the sandbox.

Orchestration

Complex tools orchestration

Coordinate multi-step workflows across APIs, databases, and services.

See example
Memory

Durable state across conversations

Remember user context and progress from past interactions.

See example
Planning

Tasks & goals

Break down complex objectives into trackable, executable steps.

See example
Discovery

Explore side effects & prerequisites

Discover dependencies and impacts before taking action.

See example
Forms

Generate forms for user input

Create interactive forms to gather missing information.

See example
Safety

Dry run before execution

Test actions safely without making real changes.

See example
Reification

Crystallize reusable workflows

Convert successful patterns into reusable functions.

See example
Select an example above to see it in action

Use Cases

Concrete use cases

Real problems solved by sandboxed agents in production

Support

Customer Support Agent

Handles customer issues end-to-end: collects missing details, performs actions in your systems, and follows up automatically.

  • Picks up where it left off, even days later
  • Checks past solutions before trying again
  • Only does what it's allowed to do

Gives you full visibility into every action.

Orchestration

API Orchestrator

Figures out what needs to be done across several APIs, stitches together multi-step operations, handles errors, and retries intelligently.

  • Plans multi-step workflows before executing
  • Saves progress if something fails
  • Shows exactly what it did and why

Provides full audit trails for compliance.

Onboarding

Onboarding Agent

Guides new users through multi-step onboarding, gathers needed information, takes actions in your product, and personalizes the experience.

  • Remembers each user's preferences
  • Remembers where you left off
  • Changes the flow based on what you need

Provides personalized experiences that scale.

Operations

Internal Operations Agent

Automates internal workflows like provisioning accounts, updating records, preparing documents, and coordinating multi-team processes.

  • Works on processes that take days or weeks
  • Stops to ask questions, then continues
  • See exactly what happened at each step

Provides complete audit trails for compliance.

Pricing

Pay per usage

Only active sandbox time is billed. Idle sandboxes are free.

A sandbox maps to a workflow, customer, or long-running task.

$0.50 / sandbox / hour

Billed per minute. Only pay when sandboxes are running.

  • Unlimited agents
  • Unlimited users
  • Unlimited sandboxes
  • Unlimited tools

Everything included

All plans include the core platform features you need.

  • AI SDK + MCP output
  • Human-Agent interactions
  • Per-user isolation
  • 7 days logging
  • Persistent state

Talk to us

Tell us about your stack and workloads. We'll map the sandbox, pricing, and rollout plan to your product.

PlatformWhite labelCustom agent

Tell us about your agentic needs

We respond with a sandbox blueprint, pricing, and rollout steps.

We reply with a sandbox blueprint, pricing, and rollout steps.