MiWork
KOEN
AI GATEWAY & GOVERNANCE

Manage your organization's AI use
in one place

MiWork Gate applies budgets, permissions and personal data policies where coding tools, AI agents and internal apps call AI models. Review cache savings and usage directly in the web console.

Contact sales Explore features
OpenAI · Anthropic · OpenAI-compatible Budget & Pricing Response Cache Privacy Policy Audit Log
Operations · This month Policies active
Charges$1,284Budget deduction basis
Provider cost$1,070Price at request time
Margin$214Charges − Cost
Organization budgets
Business support46%
Research & development84%
Sales28%
Recent security events
MaskedTwo email addresses masked before forwarding
BlockedRequest contained an API-key-like secret
WarningSuspected prompt attack pattern logged

More AI tools. Fragmented governance.

Route model requests through one gateway to apply consistent access rules and logging.

THE CHALLENGE THE MIWORK GATE APPROACH
Separate AI keys by team make usage hard to trackSee usage and cost by user and team key in one console
Costs become clear only after budgets are exceededCheck remaining budgets before forwarding and reject requests beyond limits
Savings are hard to quantifyReview cached responses and savings by period and model
Sensitive work still goes to external modelsRestrict teams to internal LLMs and mask or block personal data
Provider costs and internal charges are managed manuallyManage model costs, charge rates and margins in one place
Configuration changes are hard to traceAudit changes to policies, budgets and keys, plus prompt access
CAPABILITIES

Key features

Connect in a familiar way

Change the endpoint and key in your existing tools. Connect OpenAI-compatible models and internal LLMs in the same way; confirm support for advanced features.

Organization and team budgets

Set budgets for organizations, teams, users and keys, with periodic replenishment. Block over-budget requests before they reach the model.

Pricing and margins by model

Set provider costs and internal charge rates by model to review margins. Historical requests retain the rates that applied at the time.

Personal data and security policies

Mask patterns resembling email addresses, phone numbers, Korean resident registration numbers and card numbers, and block requests containing secrets. Organizations and teams cannot weaken higher-level security rules.

Optional prompt logging

Prompt body storage is off by default. When enabled, personal data and secrets are removed before encrypted storage; only separately authorized people can view the content.

Model connections and test calls

Connect models in the console and compare responses and costs side by side. Test calls follow the same policies and count toward usage.

Failover and load distribution

Distribute requests across multiple connections for a model. Failover is off by default; when enabled, it retries another connection on failure. Connections do not switch after a response begins.

IP access restrictions

Set allowed addresses separately for administration and model requests. Other addresses are rejected before requests reach the model.

Notifications and key lifecycle

Console alerts warn when budgets run low. Keys can expire; rotating a key immediately revokes the previous one.

Need to review policies and a configuration for your organization? Talk to our team →

RESPONSE CACHE

See your AI savings,
not just your spend

When a previous response can be reused, no model call is made, no charge is incurred and the budget is unchanged. Savings accumulate in the console.

  • —Compare savings and cached responses by period and model
  • —Savings use the actual cost of the original response
  • —Cached responses are encrypted and not shared across users
Operations · This month USD
Cache savings$318.40Amount not charged
Cached responses4,126 responsesNo model call or budget deduction
Costs by modelCached responsesSavings
Work assistantExternal models2,840$241.10
Code reviewExternal models1,032$71.85
Document summaryInternal LLMs254$5.45
Allowed models by teamWithin organization policy
HRHigh security
Internal LLMs
Personal-data-heavy work · No external models
DevelopmentStandard
Internal LLMsExternal coding model
External models for code; internal LLMs for everyday questions
MarketingStandard
Internal LLMsExternal general-purpose model
Start drafts with a lower-cost internal LLM
Illustrative configuration
MODEL POLICY

Generative AI usage security —
personal data policies and team model access

Connect internal LLMs through the same gateway as external models. Choose allowed models by team according to security and cost requirements.

  • —Restrict teams handling sensitive data to internal LLMs
  • —Use internal LLMs for everyday work and external models for demanding tasks
  • —Set allowed models only within higher-level permissions
HOW IT WORKS

Before a request reaches the model

STEP 1VerifyVerify the key, permissions and allowed model.
STEP 2InspectMask personal data and block requests containing secrets.
STEP 3ReuseReturn an available cached response without a charge.
STEP 4RouteCheck the remaining budget, then forward to an allowed model.
STEP 5SettleCalculate and record cost from actual usage.

Requests blocked by policy or completed without usage are not charged. Requests whose usage cannot be fully verified are kept for review rather than recorded as zero-cost.

ADMIN CONSOLE

An admin console scoped to your role
Admin console

Manage operations, requests, budgets, models, pricing, policies and audit logs in one web console. Use local accounts or Microsoft Entra ID for login.

Signing in does not grant administration rights. Viewing prompt bodies requires separate authorization, independent of role.

System administratorAll organizations · Models, pricing, global policies and admin accounts
Organization administratorOwn organization · Teams, budgets, users, keys and policies
Team administratorAssigned teams · Users, keys and team policies
Audit viewerAssigned organizations & teams · Read-only; no configuration changes
USE CASES

Common use cases

  • Consistent governance across internal AI toolsManage tool permissions and policies in one place.
  • Report AI cost savingsReview cache savings and model costs by period.
  • Allocate AI costs by departmentUse organization and team budgets with model-level costs and charges.
  • Work involving sensitive informationLimit the team to internal LLMs and mask or block personal data.
DEPLOYMENT & OPERATIONS

Deployment and operations

  • —Install and operate on your organization's servers
  • —Separate model request handling from administration
  • —Operate a separate Gate instance for each integrated application
  • —Encrypt model access keys, stored prompts and cached responses
  • —Keep cost records intact and track corrections separately

We recommend restricting administration to internal networks. The detailed configuration is determined after reviewing your deployment environment.

Frequently asked questions

Which AI models can I connect?+

Connect OpenAI and Anthropic models, OpenAI-compatible models such as DeepSeek, GLM and Qwen, and internal LLMs. Some features, such as provider-side paid search tools, are not supported. Confirm the models and features you plan to use in your environment.

What happens when a budget is exceeded?+

Remaining budget is checked before forwarding. Requests that cannot be covered are rejected without reaching the model. Conservative cost estimates mean even short requests may be rejected when little budget remains. Cached responses cost nothing and may still be returned after the budget is exhausted.

Can it filter out all personal data?+

Pattern-based checks cannot guarantee detection of all personal data or attacks. Uninspectable content such as images and files is forwarded and logged by default; enabling stricter inspection blocks such requests. Model responses are passed through unchanged.

Are prompt contents stored?+

Not by default. If body storage is enabled, personal data and secrets are removed before encrypted storage and deletion after the configured period. Usage and cost logs are separate. Response caching stores encrypted model responses for a set time, not the original request text.

Where can I see cache savings?+

The console's operations view shows savings, cached response counts and provider cost savings by period. The model cost table shows cached responses and savings per model. Savings depend on how many requests can reuse a response in your workflow.

Which requests can use the cache?+

The same user can reuse a response within its retention period for an identical request using the same model and options. This is not similarity-based search. Requests with files, images or continued conversations are not cached. You can bypass the cache when a fresh answer is needed.

Can I connect an internally operated LLM?+

Yes. Register internal LLM servers supporting OpenAI or Anthropic request formats alongside external models. Administrators predefine allowed server addresses. Gate does not automatically change models based on request content; team model permissions control the available scope.

What happens if an AI provider fails?+

By default, a request is sent once. If an administrator enables failover for the model, a connection failure triggers a retry through another connection. No switch occurs once response delivery begins. Retry billing rules are explained during deployment planning.

Lower AI costs. Establish clear governance.

Tell us about your AI tools, models and internal LLMs. We will review a MiWork Gate configuration that fits your organization.

Discuss MiWork Gate Explore other solutions