Resources
Company News
5 min read

Asyrx for Claude: the API, Claude.ai, Claude Code, Cowork and Microsoft 365

Claude now acts in terminals, documents, spreadsheets and business systems. Asyrx gives each consequential proposed action an Action Risk Score and Blast Radius Score before it runs.

Asyrx Research

Cyan AI assistant node linked to terminal, document and spreadsheet panels through a shield checkpoint

Claude is no longer only a chat window. The same model family now writes and runs code in a developer's terminal, works through multi-step tasks on a knowledge worker's desktop, drafts and edits files inside Microsoft 365, and powers custom agents that enterprises build directly on the Claude API. Every one of those surfaces can propose actions with real consequences: a deleted branch, a bulk email, a changed permission, an exported customer list.

Asyrx is Claude's flagship integration because it answers one question for all of them in the same way: how consequential is this specific proposed action, right now, in this context? Asyrx returns decision-support risk intelligence. The enterprise, its policies and its people keep authorization and execution control.

How the connection works

There are two connection paths. Claude.ai, Claude Code and Claude Cowork connect to the Asyrx remote MCP server, protected by OAuth. A user signs in with an approved Asyrx account, reviews the requested access on a consent screen, and Claude can then call Asyrx tools such as listing available sandboxes and assessing a proposed action. Access is scoped to the user's own organization memberships and only works for active, approved Enterprise Sandbox access.

Custom agents built on the Claude API call the Asyrx assessment API directly, with one request before a consequential tool call executes. That path supports persistence, idempotency keys, webhooks and the full audit record. Claude for Microsoft 365 deployments follow whichever path the organization uses to extend Claude with tools: a connector to the remote MCP server, or an API call from the orchestration layer that sits between Claude and Microsoft Graph.

Example use cases by Claude surface
SurfaceExample proposed actionWhat Asyrx surfaces
Claude API (custom agents)Issue a refund above the usual range to a new accountElevated Action Risk Score, financial reason codes, recommended human review
Claude.aiAssess a draft bulk update before a user pastes it into a CRMBlast Radius Score reflecting record count and downstream systems
Claude CodeForce-push to a protected branch or change CI secretsPrivilege and integrity reason codes, high blast radius across deployments
Claude CoworkMove or delete a large folder of shared files during a cleanup taskDestructive-action risk factors and data-scope context
Claude for Microsoft 365Share a workbook containing customer data with an external domainExternal destination and sensitive-data reason codes

Claude Code: assessment inside the developer loop

Developers adopt Claude Code because it moves fast. The goal is not to slow it down for routine edits, but to give consequential operations a second look. A team can ask Claude Code to assess a proposed deployment, migration or permissions change through Asyrx before running it, and receive the score, the reason codes and advisory recommended controls in the same session. Low-band actions continue as normal; higher-band actions are surfaced to the developer with a clear explanation of why.

Cowork and Microsoft 365: actions on everyday work

Desktop and productivity agents act on files, calendars, mailboxes and shared drives. Individually small actions can add up: one rename is routine, while ten thousand renames or an external share of a finance folder is not. Asyrx's Blast Radius Score is designed for exactly this distinction. It estimates how far the effects of an action could spread, separately from how risky the action is in itself, so reviewers do not confuse a risky-but-contained action with a routine-but-wide one.

Asyrx score bands applied to every Claude surface
Low (0–39)0–39
Moderate (40–64)40–64
High (65–84)65–84
Critical (85–100)85–100

Conceptual illustration of the published Asyrx score bands. It does not represent measured customer data.

What stays the same

Asyrx does not replace Anthropic's own safeguards, Claude's permission prompts, or your identity, authentication, authorization and RBAC systems. Those layers decide who may act and what tools are available. Asyrx adds action-level context on top: a consistent score from the same canonical engine, whichever Claude surface proposed the action. Sandbox assessments made through MCP are non-persisting previews, so teams can evaluate safely before wiring the production API path.

Getting started with Asyrx for Claude

  • Request Enterprise Sandbox access and complete approval for your organization.
  • Add the Asyrx remote MCP server as a custom connector in Claude.ai, Claude Code or Claude Cowork.
  • Sign in and approve access on the Asyrx consent screen.
  • Ask Claude to assess a realistic consequential action and review the reason codes.
  • For Claude API agents, add one assessment call before consequential tool calls.
  • Decide in advance how your systems respond to each score band.

Frequently asked questions

Which Claude products work with Asyrx?
Claude.ai, Claude Code and Claude Cowork connect through the Asyrx remote MCP server. Custom agents on the Claude API call the Asyrx assessment API directly. Claude for Microsoft 365 deployments can use either path, depending on how the organization extends Claude with tools.
Does Asyrx approve or block Claude's actions?
No. Asyrx returns decision-support risk intelligence: an Action Risk Score, a Blast Radius Score, confidence, reason codes and advisory recommended controls. Your systems, policies and people keep authorization and execution control.
How does Claude sign in to Asyrx?
Through OAuth. A user signs in with an approved Asyrx account and approves access on a consent screen. Access is limited to that user's organization memberships and requires active, approved Enterprise Sandbox access.
Are MCP assessments stored?
Sandbox assessments made through MCP are non-persisting previews from the canonical Asyrx engine. Persistence, idempotency and webhooks are provided by the Asyrx API path.
Does Asyrx replace Claude's built-in permission prompts?
No. Claude's permission prompts and Anthropic's safeguards remain in place. Asyrx adds a consistent, explainable assessment of how consequential a specific proposed action is.
Is Asyrx affiliated with Anthropic?
Asyrx is operated independently by Thought Media™. It connects to Claude using Anthropic's published support for remote MCP servers and the Claude API.
ClaudeMCPIntegrationsAction Risk