Skip to main content Announcing Tool Policies: AI agent permissions down to the field Read the announcement

AI agent control. Control every agent your company runs

StackOne Agent Control checks every agent action before it runs, on any device, model or cloud.

Reading the Azure CLI token cache was blocked in Claude Code

Today 14:25 · noah-mbp · macOS 15.4

Blocked

deny-azure-credential-reads-command Decided on the device

The call Bash cat ~/.azure/msal_token_cache.json
  • Person Noah Patel, from Okta
  • AI tool Claude Code, hook in place
  • Rule deny-azure-credential-reads-command
  • Mode Enforcing
Chain unbroken

Agents now write to HR, finance and sales systems and run on every laptop.

They inherit everything the person can access, and nobody checks the action before it runs.

Where AI agents run. One policy, checked wherever the agent runs

On a Mac, on Windows, in the browser, in a sealed container or in the code your team ships, each action is checked against your policy before it runs, and the person sees what happened and why.

~/billing · claude

>Deploy the billing service to staging

Bash(security find-generic-password -s aws-vault -w)

Blocked by StackOne Agent Control. Passwords in the macOS Keychain stay out of reach of AI tools (policy: deny-keychain-reads-command). Ask your administrator if you need access

The Keychain is blocked by policy, so I'll deploy through the CI pipeline's role instead.

StackOne Agent Control

This computer

What protects the AI tools on this computer, and what it has stopped.

Claude Code is protected. Its calls are checked against your organisation's policy.

AI tools on this computer
ToolStatusLast checked a call
Claude Code Protected Just now
Cursor Protected 2 h ago
Codex Protected Yesterday

Recent blocks Tool calls that policy stopped on this computer, newest first

Just now Reading the macOS Keychain isn't allowedClaude Code deny-keychain-reads-command
Claude Code on a Mac. The Keychain read stops in the terminal, and the Agent Control app shows the person what was blocked and why.

AI agent control architecture. One control layer across your whole estate

Agent Control decides on every action, and the Tool Gateway connects agents to your systems. Both run in our cloud or yours, next to your clouds, on-prem and laptops.

Your estatePublic clouds, on-prem, sovereign, desktops

Your agents

StackOne

On each laptop

In each container

In your code

Your MDM

Agent Control

Tool Gateway

Your identity provider

Your EDR

Your SIEM

Your systems

Coding and chat agents, cloud agents, agents you build

  • Claude Code
  • Codex
  • Cursor
  • Copilot CLI

Hooks and process guard

  • macOS
  • Windows

One-line install

  • Docker

SDK hooks

  • Policies
  • Inventory
  • Decisions

Connectors and remote MCP servers

  • Jamf
  • Intune
  • Okta
  • Entra ID
  • CrowdStrike
  • SentinelOne
  • Splunk
  • Datadog

SaaS and internal APIs

  • Workday
  • Salesforce
  • SAP
  • ServiceNow

Installs the app

Signed policies out

Records home, kept offline

Tool calls

Decides each call

Checked calls

People and groups (OIDC)

Signals out (tampered hook)

Every decision (Logs API)

governed callasync·Logos show examples of each kind of tool.

  1. Your agentsCoding and chat agents, cloud agents, agents you build Claude CodeCodexCursorCopilot CLI
    • On each laptop Hooks and process guard macOSWindows
    • In each container One-line install Docker
    • In your code SDK hooks
  2. Signed policies out
  3. Records home, kept offline
  4. Tool calls
  5. StackOne Agent ControlPolicies, inventory, decisions Decides each call Tool GatewayConnectors and remote MCP servers
  6. Checked calls
  7. Your systemsSaaS and internal APIs WorkdaySalesforceSAPServiceNow
  8. Your identity provider: People and groups (OIDC) OktaEntra ID
    Your EDR: Signals out (tampered hook) CrowdStrikeSentinelOne
    Your SIEM: Every decision (Logs API) SplunkDatadog
    Your MDM: Installs the app JamfIntune
  9. Your estate: Public clouds, on-prem, sovereign, desktops. Logos show examples of each kind of tool.

AI agent control console. Every action, agent and device, in one console

We check the agent, the person, the data and the action before anything runs.

Decide on every action

approval-for-writes

Published · version 3

Mode: Ask
What the person is told

Tools that create, change or delete records need a person to approve them.

Held 4 writes today for an HR partner

StackOne14:31

Alex Morgan asked Claude to change a pay rate in Workday. It needs an HR partner.

Worker
Priya Shah
Base pay
£48,500 → £52,000
Effective
1 Dec
Lapses
in 24 hours

Approving lets exactly these values through, once. Nobody can approve their own request.

  • Watch before you block. Start new rules in Monitor only, so you see what they would block before they block anything.
  • Approvals go to Slack, Teams, email or the StackOne app, and cover only the values that were approved.
  • Each call is also checked for prompt injection and personal data, by our own models.
  • Every call is logged: decision and reason, request logs, session timeline, SIEM export.

See every agent

AI inventory

Every agent, MCP server and AI app, with its owner and where it was found

AI inventory
App or agentFound byOwnerState
Claude CodeAI tool Device app Noah Patel Enforced
Microsoft CopilotAI app Entra ID Lena Ortiz Enforced
CursorAI tool Device app Ravi Menon Monitored
github MCP serverMCP server Device app Noah Patel Enforced
ChatGPT, personalAI app Web gateway logs Maya Chen Discovered
Gemini CLIAI tool EDR inventory Jordan Blake Discovered
  • Import. From your identity provider's grants, AI vendor audit logs, EDR and MDM inventory, web gateway logs, and the Agent Control app on each device.
  • One inventory. Every agent, MCP server and AI app in one list, with its owner and what it can reach, marked Enforced, Monitored or Discovered.
  • One allowlist, enforced. Allow only the AI clients, models and tools you approve, and enforce it on the StackOne route, in managed device settings, through your identity provider and in AI vendor admin consoles, even where an agent never touches StackOne.
  • Any AI tool your policy hasn't approved shows up as shadow AI, on the laptop and in your vendor logs.

Every device, checked

Devices

Protected87 / 8
AI tools checked16
On the newest version8 / 8
Devices
DevicePersonStateAI toolsLast seen
noah-mbp Noah Patel ProtectedHook changed
lena-mbp Lena Ortiz Protected
ravi-win Ravi Menon Protected
maya-mbp Maya Chen Protected
jordan-win Jordan Blake Protected

Agent Control

Protected Needs attention
  • Claude Code ProtectedRepairing
  • Cursor Protected
  • Codex Protected

The protection hook was changed Repaired in 4 s, and logged Blocked 2 min ago: reading credentials from the shell

Policy 2 · synced

  • Hook settings are repaired within 5 seconds if anyone changes them, and the change is logged.
  • The process guard catches an agent started with its hooks switched off.
  • Each person sees whether they're protected and what was blocked, from the menu bar or system tray.
  • Time a device spent unprotected, with the service stopped or StackOne out of reach, is recorded with a start and an end.
  • Each person sees their own agents and usage, and can add policies that only narrow what the organisation allows.

Kill switch: stop every AI tool on every device. One button in the console stops them across the organisation, and records who stopped them and why.

Org-wide. Each device stops at its next check-in, within a minute, and every refused call is still recorded.

AI tools are stopped on every device.

Started by Noah Patel at 14:31: a build runner's token may have leaked.

AI tools can run again. Devices pick it up within a minute.

Devices

All 6 devices protected

  • noah-mbp Noah Patel · macOS Protected Not stopped yet AI tools stopped
  • ravi-win Ravi Menon · Windows Protected Not stopped yet AI tools stopped
  • lena-mbp Lena Ortiz · macOS Protected Not stopped yet AI tools stopped
  • build-runner-07 Noah Patel · Linux container Protected Not stopped yet AI tools stopped
  • maya-mbp Maya Chen · macOS Protected Not stopped yet AI tools stopped
  • jordan-win Jordan Blake · Windows Protected Not stopped yet AI tools stopped

AI security models. Built AI-first: our own models and our own connectors

Most vendors bolted agents onto existing products. We built the Tool Gateway, Agent Control and the company itself for the AI era. Off-the-shelf security tools and generic models make bad control decisions and get in people's way, so we trained and fine-tuned our own.

StackOne Defender, the prompt injection model in Agent Control, scores 0.94 F1 on 1,071 InjecAgent tool responses, against 0.79 for ProtectAI DeBERTa-v3, and checks each call in about 2.5 ms on a CPU from a 22 MB model. StackOne PII Redaction scores 97.9% span F1 on 247 real connector field values, against 89.6% for GLiNER multi PII v1, with half the false positives.

Injections in tool results, F1
  1. StackOne Defender 0.94+20% StackOne Defender StackOne's own prompt injection model (version 0.7.0), 22 MB, about 2.5 ms per check on a CPU.F1 0.94 on 1,071 tool responses.20% higher than ProtectAI DeBERTa-v3, the next best.
  2. ProtectAI DeBERTa-v3 0.79 ProtectAI DeBERTa-v3 ProtectAI's open-source prompt injection classifier.F1 0.79 on 1,071 tool responses.StackOne scores 20% higher.
  3. NemoGuard JailbreakDetect 0.31 NemoGuard JailbreakDetect NVIDIA's jailbreak detection model from NeMo Guardrails.F1 0.31 on 1,071 tool responses.StackOne scores 3x higher.
  4. Meta Prompt Guard v2 0.04 Meta Prompt Guard v2 Meta's open prompt attack classifier.F1 0.04 on 1,071 tool responses.StackOne scores 23x higher.

How it was measuredInjecAgent: 1,071 tool responses, 1,054 of them attacks. F1 at the default threshold, Defender 0.7.0. Defense benchmark on defender.stackone.com (opens in a new tab)

  • Injections in tool results

    0.94F1 on 1,071 tool responses carrying hidden instructions

    20% higher F1 than the next best

    F1

    Next best: ProtectAI DeBERTa-v3, 0.79Defense benchmark on defender.stackone.com (opens in a new tab)

  • PII in connector data

    97.9%span F1 on 247 real connector field values

    9% higher F1, half the false positives of the next best

    Span F1

    Next best: GLiNER multi PII v1, 89.6%PII Redaction benchmark

  • Runs on a laptop CPU

    22 MBprompt injection model, 2.5 ms per check on a CPU

    26× smaller than the next smallest, and no GPU

    Model size, smaller is better

    Next smallest: NemoGuard JailbreakDetect, 563 MB on a GPUDefense benchmark on defender.stackone.com (opens in a new tab)

Smaller and faster sits bottom left Size and time per check, ours against the next best
StackOne DefenderStackOne's prompt injection model.22 MB on disk.About 2.5 ms per check on a CPU.26× smaller and 2.4× faster than NemoGuard, with no GPU. NVIDIA NemoGuardNVIDIA's jailbreak detection model from NeMo Guardrails.563 MB on disk.About 5.9 ms per check on a GPU.26× the size of StackOne Defender, 2.4× slower. StackOne PIIStackOne's PII detection model.24 MB on disk, 205 MB of memory running.About 1 ms per field on a CPU.19× smaller and 1.8× faster than Presidio. PresidioOpen-source PII detector.447 MB on disk, 1.0 GB of memory running.About 1.75 ms per field on a CPU.19× the size of StackOne PII, 1.8× slower. GLiNER multi PII v1Open multilingual PII model.1.2 GB on disk, 2.7 GB of memory running.About 146.43 ms per field on a CPU.48× the size of StackOne PII, 146× slower. OpenAI Privacy FilterOpenAI's open PII model.2.8 GB on disk, 4.3 GB of memory running.About 384.85 ms per field on a CPU.117× the size of StackOne PII, 385× slower. NVIDIA GLiNER-PIINVIDIA's open PII model, trained on Nemotron-PII.1.8 GB on disk, 3.9 GB of memory running.About 496.7 ms per field on a CPU.75× the size of StackOne PII, 497× slower. Ettin-68MOpen model fine-tuned on Nemotron-PII.141 MB on disk, 494 MB of memory running.About 14.6 ms per field on a CPU.5.9× the size of StackOne PII, 15× slower. Isotonic DeBERTa-v3Open model trained on Ai4Privacy data, the publisher of OpenPII.747 MB on disk, 877 MB of memory running.About 55.97 ms per field on a CPU.31× the size of StackOne PII, 56× slower. Piiranha mDeBERTaOpen multilingual model trained on Ai4Privacy data, the publisher of OpenPII.1.1 GB on disk, 997 MB of memory running.About 55.24 ms per field on a CPU.47× the size of StackOne PII, 55× slower.
  • StackOne Defender against NVIDIA NemoGuard

    26× smaller, 2.4× faster, no GPU

  • StackOne PII against Presidio

    19× smaller, 1.8× faster

How it was measuredPII times are the median per connector field on 4 vCPU through one harness. Defender (0.7.0) and NemoGuard are from defender.stackone.com: Defender on a CPU, NemoGuard on an A10G GPU. Hover a point for its numbers

Test us: bring your hardest cases and we'll run our models on them.

640+ connectors we build, so rules understand each field.

Book Demo

AI agent identity. Agent identity answers two questions on every call

Who the agent is acting for, and which agent it is. The call can only do what the person can do and the agent is allowed to do.

When it acts for a person

The call can use only what the person can do, from your identity provider, and what the agent's mandate allows.

When it acts on its own

Each agent signs in as itself, through Entra Agent ID or its own StackOne key, with a named owner, an expiry date and a switch to turn it off.

Mandates only ever narrow access: an agent gets a restricted profile of what the person can do, never more.

Weekly payroll check

Agent identity

Owner
Lena Ortiz, Finance manager
Signs in with
Workload identity, no stored secret
What Lena Ortiz can reach, and what Weekly payroll check may do with it
System Lena, from Okta The agent's mandate
Workday Read and change payrollRemoved by Okta Read payroll runsPaused: Lena can't reach this
SAP S/4HANA Read and approve invoicesRemoved by Okta Read invoicesPaused: Lena can't reach this
Salesforce Read accountsRead accounts No access
ServiceNow Open ticketsRemoved by Okta No access
Slack Post in finance channelsPost in finance channels No access

The mandate is checked against Lena's access on every call.

Explore SSO and SCIM for AI agents

Production AI agents. Same rules for the agents you build

Production agents run on servers, not laptops. Inside the agent, the SDK decides which tools it gets and checks every call. Around it, the sandbox decides what it can reach.

Inside the agent, with the SDK

  • Which tools it loadsOnly the tools its mandate allows. The rest never reach the model.

  • Every call and its inputsField and value rules, and a person's approval for the risky ones.

  • What comes backPersonal data masked and injections caught before the model reads them.

  • Prompts and repliesChecked in the code you ship: LangGraph, the Claude Agent SDK and the AI SDK.

Around the agent, in the sandbox

  • NetworkOut only through the StackOne gateway, with rules on each request.

  • Files and foldersSecret paths denied, compiled into the sandbox settings.

  • Commands and packagesUnknown registries refused. One line in your Dockerfile installs it.

Across both

  • Who it isSigns in as itself with no stored secret, and acts for a person only within that person's access.

  • Where calls runSelf-host the gateway beside your agents, so calls to internal APIs stay in your network.

  • Every decision on recordSent to your SIEM through the Logs API.

AI agent control vs EDR, DLP and SIEM. We decide what agents do beside the security tools you already run

None of these tools was built to read what an agent's action changes and whether it should take the action. StackOne does, and shares each decision with them.

Agent Control What agents do

MDM puts software on the laptop

Where we overlap
Your MDM installs our app and hooks. We only manage agent hooks and settings.
Stays with your MDM
Apps, OS updates and device policy. We push nothing unrelated to agents.
For example Examples of tools we work beside, not partnerships.
  • Jamf
  • Microsoft Intune

EDR watches devices and networks

Where we overlap
We check network traffic agents are likely sending, and send you signals like a tampered hook or a credential read.
Stays with your EDR
Antivirus, malware detection and forensics.
For example Examples of tools we work beside, not partnerships.
  • CrowdStrike
  • SentinelOne

Identity knows who is acting

Where we overlap
We identify each agent and tie it to a person. Some identity providers, like Okta, are adding agent identity too.
Stays with your identity provider
Sign-in, MFA, people and groups. We read them from yours.
For example Examples of tools we work beside, not partnerships.
  • Okta
  • Microsoft Entra ID
  • Google

DLP stops data leaving

Where we overlap
We redact and block personal data in agent actions, by policy, and log every case.
Stays with your DLP
Email exfiltration, USB, printing and file scanning.
For example Examples of tools we work beside, not partnerships.
  • Microsoft Purview
  • Google

Your SIEM sees events, after the fact

Where we overlap
Every decision goes to it through our Logs API.
For example Examples of tools we work beside, not partnerships.
  • Splunk
  • Datadog
  • Elastic
Agent Control beside MDM, EDR, identity, DLP and SIEM
Tool Where Agent Control overlaps Stays with that tool
MDM Your MDM installs our app and hooks. We only manage agent hooks and settings. Apps, OS updates and device policy. We push nothing unrelated to agents.
EDR We check network traffic agents are likely sending, and send you signals like a tampered hook or a credential read. Antivirus, malware detection and forensics.
Identity We identify each agent and tie it to a person. Some identity providers, like Okta, are adding agent identity too. Sign-in, MFA, people and groups. We read them from yours.
DLP We redact and block personal data in agent actions, by policy, and log every case. Email exfiltration, USB, printing and file scanning.
SIEM Every decision goes to it through our Logs API. Every event from every tool, after the fact.

On-device enforcement. Decides on every device, syncs with the deployment you choose

Each laptop runs the policy engine and our small models itself, so checks keep working offline and most never leave the device. Larger models in the deployment you pick take the calls your rules send them.

On every device, always on

  • Policy engineUnder 1 ms per policy check, in 13 MiB of memory

  • Prompt injection model2.5 ms per check on a CPU, from a 22 MB model

  • PII modelAbout 1 ms per field on a CPU, from a 24 MB model

  • Alignment model, in developmentUnder 1% added to an agent's turn on an M4 Pro's GPU

  • Signed policies, cachedThe last rules pushed keep applying with no connection

Connected to your deployment

Signed policies are current, and records go up as they happen.

0 records waiting

Works offlineNo call out to decide

In the deployment you choose

Shared cloudDedicated accountSelf-hosted

  • Larger modelsCalled when a rule asks for a deeper look, like a write to a payroll system.

  • Console and policiesWhere rules are written, signed and pushed to every device.

  • Every decisionKept and sent on to your SIEM.

01

Shared cloud

StackOne runs outside your perimeter, in StackOne cloud, shared with other customers.

Encryption keyOurs or yours

StackOne hosts and operates the service.

Choose a region for each project EU: Ireland, Belgium US: Northern Virginia Other regions on demand
02

Dedicated account

StackOne runs outside your perimeter, in a StackOne cloud account that serves only your organization.

Encryption keyOurs or yours

An environment dedicated to your organization.

Agree region and operating arrangements
03

Self-hosted

StackOne runs inside your perimeter, on AWS, Google Cloud, Microsoft Azure, Oracle Cloud, Rackspace or any Kubernetes-compatible environment.

Encryption keyYours

Your team operates StackOne in your environment.

Separate license and support agreement

The same checks in every option. You choose where data, keys and logs live.

Explore deployment and self-hosting

Customers. Security-first teams trust us with their most sensitive systems

From published StackOne case studies, on the platform Agent Control runs on.

Medical research laboratory Healthcare · Customer story
"The granular access at the account level for a connector is a huge help. So now I feel confident that we can create connectors to everything."
IT leaderUS medical research organisation
Read the US medical research organisation story
Signing dayFirst live agent query
"As a company rooted in security assurance, we needed a partner sharing our priorities. StackOne's approach gave us confidence to scale without compromise."
Faraz YaghoutiSenior Director of Product, Drata
Read the Drata story

SOC 2 Type II

Request the report, and review subprocessors and security information, in our trust center.

Visit the trust center

The StackOne platform. Connect agents to your systems, and make them faster

FAQ. Questions about AI agent control

See Agent Control on your own agents

Bring the agents your teams run and your hardest cases, and we'll show each action checked, decided and logged.