Skip to main content

Your security policyapplied to every agent.

Speakeasy provides the organizational security layer agents are missing. Every action is checked against your policy with the whole session context considered. Sessions in violation are paused and escalated.

Deploy for free

Trusted by

The biggest companies
Google
Fifth Third
Mistral
Cisco
The fastest startups
PlanetScale
Vercel
Polar
MoonPay

To secure agents,
you need to understand intent.

Agent security isn't black & white. To understand what's allowed, you need to evaluate the agent's intent against the full context: the user's ask, the company's policies, the agent's plan/

The inbox takeover

An email the agent is triaging tells it to post a customer's billing record to #all-staff.

The borrowed prod token

Debugging staging, the agent finds a production admin token in config and uses it.

PII in a public issue

An investigation puts a customer's name, card, and address in a public GitHub issue.

The surprise force push

Asked to fix a failing test, the agent decides to force-push to main.

Yes to which cleanup?

Offered two cleanup plans, the user says yes, and the agent picks the destructive one.

The finance escalation

With a restricted invoice in context, a minimal summary goes to #finance-ops only.

Becoming AI-native meant that we needed an entirely new governance & security stack. You can't use old tools to cover your team's new way of working.

Shreyas Kumar

Co-founder, Fermat

Fermat

Every agent follows
company policy

Start from Speakeasy's library of policies, or write your own for the way your teams work. Speakeasy checks every agent action against them with the whole session in view, and asks the right person when an action falls outside them.

01

Capture every action

Hooks in coding agents and the MCP gateway for web apps put every tool call in the path. Roll out through Jamf, Intune, or the MDM you already run.

02

Classify the session

Classification models read the session as it unfolds: what data it has touched, whether untrusted content is steering it, and what the next action would do.

03

Decide against policy

Your organization's policy, global and per team, is applied to that context outside the agent. A prompt injection can't talk it out of the rules.

04

Escalate when needed

Allowed actions run without a prompt. The rest stop or go to an approver in Slack, and every decision lands in the audit trail.


Security for every agent.

Agents differ on two axes: whose permissions they use, and whether anyone is watching. Speakeasy applies one organizational policy in every quadrant, and sends each escalation to someone who can decide it.

Delegated
Acts for a signed-in user
Attended
Coding agents and AI assistants
Hooks put every tool call in the path. Escalations go to a peer, the team, or security, as the policy sets.
  • Claude Code
  • Claude
  • ChatGPT
  • Codex
  • Cursor
  • GitHub Copilot
  • OpenCode
  • Gemini
  • Mistral
  • Glean
Unattended
Personal assistants working in the background
The same policy runs with nobody at the keyboard. A blocked action waits in Slack for its approver, and the session resumes once they decide.
  • Claude Cowork
  • Devin
  • OpenClaw
  • Hermes
  • Dots
Standalone
Acts as itself
Attended
Shared agents and team bots
Policy follows what each session has read, so records one person's request pulls in stay inside the audience they belong to.
  • Slack
  • Claude Tag
  • Grok Bot
  • Muse
Unattended
Backend agents, CI/CD jobs, Kubernetes workloads
Always-on checks through the MCP gateway. Escalations route to the agent's accountable owner, and every decision lands in the audit trail.
  • Azure AI Foundry
  • GitHub Actions
  • Custom agents

The browser-based approach we relied on just doesn't work for AI-based workflows.

Thierry Dang

Security Operations

MoonPay

Proven where the stakes are highest

Security teams at payments, fintech, and infrastructure companies run the control plane in production: company-wide rollouts, tens of thousands of governed agent sessions a week, and shadow MCP usage brought into one catalog.

Read the MoonPay security story

60K

Agent sessions governed per week

MoonPay enforces policy across every agent, org-wide, as of October 2026.

100%

Employee AI rollout, governed

Fermat rolled out governed AI to the whole company.

200%

MCP usage growth under governance

PlanetScale customers are now agent-based.

Your compliance needs an AI control plane

The future of compliance depends on being able to answer three questions: what AI is in use, what can it access, and who approved each action. Speakeasy's control plane produces the answers.

Speakeasy is SOC 2 Type II, ISO 27001 and CSA STAR AI certified, GDPR compliant and HIPAA ready.

EU AI Act

Logging, human oversight, and risk controls for AI systems in production. The catalog, approval flow, and audit trail supply the operational evidence.

ISO 42001

The AI management system standard. Governed catalogs, scoped access, and decision records map to the controls an auditor asks to see.

NIST AI RMF

Govern, map, measure, manage. Fleet-wide visibility and runtime enforcement cover the functions that need more than policy documents.

Compliance at Speakeasy

The control plane meets the standards it helps you enforce. Speakeasy is SOC 2 Type II audited, ISO 27001 certified, and GDPR compliant, with reports available through the trust center.

AICPA SOC for Service OrganizationsISO 27001 Information Security Management CertifiedThe Linux FoundationCloud Security AllianceAgentic AI Foundation
Visit trust center

Questions

How is this different from Claude's auto mode?
Auto mode decides which actions are safe based on what the user is trying to do, and it is good at that. It has no view of your company's rules, so an agent can follow the user's request exactly and still post customer details to a public repo. Speakeasy applies your organization's policy to every agent action, and routes blocked actions to an approver you choose instead of the user who asked.
How is this different from DLP?
DLP scans each message for sensitive patterns on its own. Agents handle sensitive data constantly, so that approach floods security teams with findings on ordinary work. Speakeasy tracks what a session has read and acts when that data is about to leave the audience it belongs to, such as a public channel, an externally shared document, or an outside recipient.
Can a user or a prompt injection change the policy?
No. Organizational policy is held outside the agent and is only presented to the classifiers that evaluate each action. Nothing in the session, whether a user message or instructions hidden in a web page or email, can edit it.
What happens when an agent is blocked?
The action stops and the agent is told why, so it can continue another way, for example by referencing a record ID instead of copying its contents. When a policy allows escalation, the request goes to the approver it names, in Slack, with the session context attached. Approvers can approve once, allow for a set time, or deny.
Who approves a blocked action?
Each policy sets its own escalation path. Low-risk policies can take a sign-off from any peer, so no single person becomes a bottleneck. Sensitive policies can require the team lead, IT, or the security team.
Does this replace role-based access control?
No, it builds on it. Whether someone can read a system at all stays in RBAC and your identity provider. Organizational policy governs what an agent does with data once it has legitimate access.
Which agents does it cover?
Coding agents such as Claude Code, Codex, and Cursor are covered through hooks deployed with your MDM. Web apps such as ChatGPT and Claude are covered through the MCP gateway, which admin consoles can make the only route to tools. Custom agents connect through the MCP gateway.
What certifications does Speakeasy hold?
Speakeasy maintains SOC 2 Type II and ISO 27001 certifications. Audit reports and compliance documentation are available through the trust center.
How do we get started?
Book a demo and our team will walk you through rollout, connecting your agents, and setting the policies and escalation paths for your organization.

AI everywhere.

Control here.