Managed confidential execution

Run agent functions with enforced limits and verifiable results

Standard runs approved functions in a confidential runtime, enforces what each operation may access and release, and returns proof your workflow verifies before the next step.

Zero-trust principles

Four protections around every operation

  • Authority outside the agent

    Owners control which functions may run, which inputs they may use, and which results may be released. Standard enforces those permissions independently of the requesting agent.

  • Restricted Wasm execution

    Functions use declared input and output interfaces with enforced resource limits. They receive no general shell, filesystem, subprocess, or network interface.

  • Confidential execution and key custody

    Functions run inside an attested TEE that protects code and inputs during execution. Supported vault and KMS connections keep non-exportable master encryption keys in your store.

  • Proof before the next step

    Your receiving integration verifies Standard Proof against the expected operation before accepting its result. Your workflow then evaluates the outcome and enforces permission for the next action.

Authorized confidential components may use temporary data keys and connector credentials for permitted processing.

How functions get data

The execution lifecycle

Define. Authorize. Execute. Release. Verify.

Explore the operation lifecycle
ATTESTED CONFIDENTIAL COMPONENTS

Example · Conversation analytics

Five stages, one operation

  1. DefineDefine conversation-categories, its conversation input, category definitions, permitted result, and execution limits.
  2. AuthorizeThe Analytics agent requests classification of batch-042. Each required owner approves its assets, the function, and permitted-result release to the Analytics integration.
  3. ExecuteAttested confidential ingress retrieves and validates batch-042. The approved classifier runs inside the confidential runtime. Raw conversations stay out of the analytics agent and coordination services.
  4. ReleaseThe permitted category counts and Standard Proof reach the Analytics integration. Successful computation alone does not authorize disclosure.
  5. VerifyThe Analytics integration verifies Standard Proof against trusted references before making the counts available to the Analytics agent.

In practice

See the protections at work

Your analytics agent requests category counts from a private conversation dataset. Standard runs the approved function and releases the permitted counts with evidence your integration verifies.

Example · Conversation analytics

This example assumes the classifier and any model it uses run inside Standard’s protected runtime.

Choose a threat to compare broad access with Standard.

An approved operation

Agent request
“Count the approved conversations by category.”
Protected input
An approved batch of 1,000 conversations from your connected source.
Approved function
Classify each conversation and total the categories.
Permitted result
CategoryConversations
Coding420
Writing310
Research190
Other80
Total1,000

Standard Proof verified. Counts available to the analytics agent.

The analytics agent receives the permitted counts. The underlying conversations and source credentials are not supplied to its environment.

Your team defines the classifier, categories, input checks, and permitted output. Standard enforces the accepted execution and release rules and returns evidence of the operation.

See which threats these protections address and where their limits are.

Threat model and AI risk mapping

Cryptographic execution evidence

Verify the execution before the next action

Verification reference

Your receiving integration checks the approved classifier, conversation batch, owner authority, runtime, and released result against references it trusts.

Verified counts become available to the analytics agent. Your reporting workflow then applies its own policy before publishing them.

Select an outcome, or alter a bound field to see what verification rejects.

Example · Conversation analytics

Standard Proof

Illustrative execution evidence

Approved function
conversation-categories · version 1
Owner authority
Classification and permitted-result release
Permitted recipient
Analytics integration
Evidence
Accepted against the receiving integration’s references.
Accepted against the receiving integration’s references.
Verify the evidence. Evaluate the result. Enforce the policy for the next action.
Analytics integrationVerifies the result before making category counts available to the analytics agent
  • Evidence: Accepted against the receiving integration’s references.
  • Function outcome: Classification completed.
  • Reporting policy: Publication permitted.
DecisionCounts ready for the reportThe receiving integration has verified the result. The analytics workflow may use the counts and publish them under its reporting policy.

Execution evidence does not certify classification accuracy or the truth of the source conversations.

Developer infrastructure

Define the function. Let your agent call it.

Publish your classification function, connect its input source, and define the permitted output. Your agent requests category counts through its existing workflow.

Example · Conversation analytics

You define the function. Standard operates the confidential runtime, including startup, execution limits, and cleanup.

  1. Deploy

    Build and test the classification function. Publish it with its input format, category definitions, execution limits, and permitted result.

    conversation-categories

  2. Authorize

    The input owner approves the selected conversation batch, the classifier, and release of the permitted result to the analytics integration.

    Approve the operation

  3. Run

    The analytics agent calls the function using the approved source reference. Standard retrieves and validates the batch, then runs the classifier inside the confidential runtime.

    Classify batch-042 and return category counts.

  4. Receive and verify

    The receiving integration verifies Standard Proof before making the counts available to the analytics agent. The reporting workflow then applies its own publication policy.

    Evidence verified · Permitted counts received

MCP — Connect your existing agent

Let your agent request the approved classification through MCP. The authorized receiving integration retrieves and verifies the permitted result.

mcp.json
{
  "mcpServers": {
    "standard": {
      "type": "http",
      "url": "https://standardruntime.com/api/mcp"
    }
  }
}

The reference includes the duplicate-ID check. Build the classifier as an approved Wasm function for protected execution.

Sources and customer vaults

Connect private sources. Keep control of your credentials.

Let agents request approved work using source and secret references. You choose the resources and permissions; Standard retrieves, validates and processes the authorized material in confidential components.

Explore sources and customer vaults
  1. 01 · Customer-controlled

    Your sources + vault

    Select a supported source and keep persistent control of keys and API secrets in your own vault or key store.

  2. 02 · Confidential ingress

    Retrieve and validate

    Check the approved connection, selection and version before retrieval. Verify the required digest and size before contribution.

  3. 03 · Operation worker

    Run the approved function

    Process validated inputs inside the attested confidential worker, with the interfaces and limits granted to the function.

  4. 04 · Authorized recipient

    Permitted result + Standard Proof

    Release the permitted result to its recipient, whose receiving integration verifies Standard Proof before accepting it.

Zero custody of your master encryption keys

For supported vault and KMS connections, your non-exportable master encryption keys remain in your store. You control their permitted use. Authorized confidential components may use temporary data keys and API credentials to retrieve and process approved inputs.

References in. Permitted results out.

Agents request work using references rather than raw source data or credentials. Only authorized recipients receive permitted results.

Bring your first agent function to Standard

Begin with one function and a clear rule for accepting its result.