AITALON AIGate
AIGate client for everyone - coming soon

AIGate AI control in browsers, IDEs and APIs

AIGate helps teams use AI safely. The client checks text before it is sent in supported chats and actions it can access in an IDE. The server checks connected model requests and agent tool calls.

MODULE_01 ACTIVE

Browser protection

The AIGate-client extension checks text before submission in supported AI chats and search forms. It blocks unsafe requests and lets the user review masked text before sending again.

WEBPREFLIGHT
CONTROL PLANE
MODULE_02 ACTIVE

AI coding IDE control

AIGate-client checks IDE actions it can access before they run. With a separate connection, model requests pass through AIGate so sensitive data can be checked and masked.

IDEHOOKS
CONTROL PLANE
MODULE_03 ACTIVE

Destructive agent action control

The server checks MCP and API calls before execution. Data deletion and other risky operations can be blocked or sent for approval, with each decision recorded in the audit log.

MCPREVIEW
CONTROL PLANE
CONTROL CIRCUIT

From a person or agent to a protected action

A person or AI agent starts the work. AIGate-client checks text before it is sent in supported chats and actions it can access in an IDE. Model and tool requests connected to Edge follow server rules.

Dotted: AIGate-client checks browser text and risky IDE commands locally Solid: configured requests through Edge and server gates
CONTROL IN BOTH DIRECTIONS AIGate-client checks actions it can access on the device. Connected requests follow server rules, and the result returns to the person or agent. POLICY ENGINE / ACTIVE
AIGate Architecture

The extension checks text before it is sent in supported chats, while the IDE client checks actions it can access before they run. Connected model and tool requests follow server rules.

Four control routes
Browser AI chat / search Extension → AIGate-client Allow / mask / block

Text is checked before it is sent in supported fields. The full site traffic does not pass through AIGate.

AI coding IDE · actions AIGate-client · hooks Allow / block

The client checks agent commands and actions it can access before they run. Coverage depends on the IDE.

AI coding IDE · model AIGate-client · bridge Edge → LLM Gate

Model requests receive server checks when the IDE is connected to Edge.

Application / AI agent API / MCP Edge → Agent Gate

Risky actions, such as deleting data, can be stopped or sent for approval.

In a browser, AIGate-client checks text only in supported fields. Address-bar queries, voice and other ways of sending data are outside this coverage. Mandatory browser control requires a managed browser.

01 / SOURCE Person or AI agent

Browser · IDE · application

AIGate-client On the device: text before sending and risky IDE commands before they run
02 / PROXY Edge

Checks access and routes connected requests

03 / SERVER POLICIES
LLM GatePrompts, sensitive data, model responses
Agent GateAgent permissions and risky actions
04 / RESOURCE Model or system

The checked result returns to the person or agent

Admin GatePolicies · approvals · limits · audit
The client checks supported browser and IDE actions. Server rules apply to requests connected to AIGate.
CL

AIGate-client

LOCAL

Checks text in supported chats and IDE actions it can access on the device. With a separate connection, model requests pass through the server. Coverage depends on the site and IDE version.

AG

Admin Gate

:8000

Django control plane: UI + API for policies, audit, billing, RBAC, API tokens, alert settings, and service instances.

LG

LLM Gate

:8001

OpenAI-compatible data plane: /v1/chat/completions, /metrics, /health. Policy checks, JWT auth, rate limits, redaction, audit.

IG

Agent Gate

:8002

Agent integration data plane: /proxy/{integration_id}/{path} and /mcp/{integration_id}. Routing and auth come from admin-gate.

CT

Controller

:8003

Recommendation service: /suggest. Analyzes events and active policy to suggest control and cost optimizations.

MS

Data masking

Policy DSL + rules engine for redacting secrets, PII, and sensitive attributes before requests leave the perimeter.

AU

Full Audit

Event pipeline and JSON logs with correlation_id: who called what and what happened. Export, alerts, and /health /ready checks.

Defense against AI agent attacks
ALLOW
Safe actions
REVIEW
Risky operations
BLOCK
Critical attacks
AG

AI AGENT ATTACKS

CRITICAL

AIGate classifies HTTP, API, and MCP actions as reads, changes, or destructive operations. Attempts to delete data, modify critical objects, or act without an applicable policy are blocked or routed for operator approval.

MCP

DANGEROUS MCP & TOOL CALLS

HIGH

Policies restrict available tools, operations, paths, and arguments. Unknown high-risk calls receive an additional check, while denied tools are not exposed to the agent.

SS

SSRF & ROUTE SPOOFING

CRITICAL

Outbound addresses are validated before execution: an agent cannot arbitrarily target the internal network, bypass an approved integration, or spoof trusted X-AIGATE headers.

BH

EMPLOYEE & AGENT BEHAVIOR

MODULE

A dedicated module attributes actions to the employee, initiator, and verified agent and highlights risk factors, rejected approvals, and destructive operations. Employees and agents are analyzed separately; raw prompts, secrets, and IP addresses are excluded from behavior aggregates.

LK

DATA LEAK

CRITICAL

Secrets, API keys, passwords, and code-exfiltration signals are detected and blocked before data reaches an external model or tool.

INJ

PROMPT INJECTION

HIGH

The system detects prompt injection and malicious instructions that attempt to make a model or agent bypass policy or invoke a denied tool.

PII

PII / PERSONAL DATA

HIGH

Personal data is masked or replaced with safe tokens before it leaves the protected perimeter.

AB

ABUSE

MEDIUM

Request, token, and outbound-call limits contain overload, uncontrolled loops, and high-volume agent actions.

Admin Dashboard
DB

Dashboard

Metrics visualization: traffic, blocks, token cost, agent activity in real time.

PL

Policies

Flexible rules per agent: allowed models, request limits, content filtering.

TK

Tokens & Keys

API key management: issue, revoke, rotate. Multi-level access rights.

RP

Reports

Automated security and usage reports. PDF, CSV export.

REAL-TIME METRICS
Requests / hour 12847
Blocked threats 342
Data masked 8923
Active agents 47
TOP-6 Business Tasks for AIGate

Where AIGate is used: data, browsers, IDEs, agent actions, spending and model routing.

LLM Cost Control

#01

Monitor and optimize language-model API spending with budget limits, token tracking, and department-level controls.

APILLM

Data checks before the LLM

#02

On routes through AIGate, policies detect secrets and personal data, then block or mask the request before it reaches the model.

PIILLM

AI agent action control

#03

MCP and API calls are checked before execution: data deletion is blocked or held for approval. The decision and outcome are recorded.

SIEMAPI

Browser AI chat protection

#04

The extension checks text before submission in supported chats and search forms. Other chats can use check-and-copy.

RBACLLM

Unified LLM Gateway

#05

One API for all models with routing, provider failover, and load balancing across your AI stack.

APILLM

AI coding IDE control

#06

AIGate-client checks IDE actions it can access before they run. With a separate connection, model requests pass through AIGate for data checks.

APIRBAC
BLBlog
FAQ

Frequently asked questions about AIGate

What is AIGate and what problem does it solve?

AIGate is a software platform for controlling language-model requests and AI-agent actions. It protects sensitive data, enforces access policies, and keeps a reviewable decision history.

Who is AIGate for and when is it needed?

AIGate is for information security, IT, architecture, and AI platform teams when employees or services use language models and AI agents and the organization needs unified control over data, actions, and costs.

How does AIGate work?

When a person or agent sends a request through AIGate, the system checks it against company rules. It can allow, block or send the request for approval. The decision is recorded.

What does an AIGate customer receive and how can the result be verified?

The customer receives unified policies, an AI usage overview, and a decision log for model requests and agent actions. Each conclusion can be checked against the event, policy version, actor, time, and final allow, review, or block decision.

How do we start an AIGate pilot and where is the data hosted?

The pilot starts with one representative LLM or AI-agent scenario, the required sources, and agreed allow, review, and block policies. AIGate can run in the cloud or on premises; routing and data-retention boundaries are fixed before launch.

How does AIGate control AI coding IDEs?

In supported IDEs, AIGate-client checks commands and agent actions it can access before they run. With a separate connection, model requests pass through AIGate so sensitive data can be checked and masked. Coverage depends on the IDE and version.

Does AIGate control AI requests in browsers?

AIGate-client checks text before it is sent in supported web chats and Google and Yandex search forms. It can block unsafe text or offer a version with sensitive data hidden. For other chats, users can check and copy text manually. Address-bar queries and voice are outside this coverage; mandatory control requires a managed browser.

Can I see an AIGate demo?

Yes. The demo shows a connected request route, a policy decision and the resulting audit event; we agree the scope beforehand.

Is there business cost control?

Yes. For connected routes, AIGate shows request volume, token usage and estimated cost based on configured model prices. Provider invoices are reconciled separately.

Does AIGate control AI agent actions?

Yes. AIGate checks connected MCP, API, and tool calls before execution. Policy can allow an action, block a destructive change, or send a risky call for approval. Coverage applies to connected routes.

How does AIGate protect data?

When a request goes through the AIGate server, secrets can be blocked and personal data masked before it reaches a model. In browsers and IDEs, checks cover only supported fields and actions; IDE model requests need a separate connection.

What server does a small AIGate installation need?

16 vCPU, 32 GB RAM and 500 GB SSD is a reference configuration, not a minimum requirement. The installer source includes a small resource preset: fewer processes and concurrent checks with the same models and thresholds. The 2 vCPU / 4 GiB RAM target still needs load qualification and is not a guaranteed minimum. Sizing depends on concurrent requests, context size, background jobs and audit retention; overload must not bypass checks.