Lawnect TokenBouncer

Control organizational AI
through policy and cost.

TokenBouncer is a system that puts company policy in control of internal AI usage and spend. Keep existing AI services and workflows as they are, connect organizational identities, and enforce limits and detailed policies by user and group — with reporting metrics, request records, and configurable failure and fallback behavior.

TokenBouncer usage and cost operations dashboard

TokenBouncer policies by user, group, and model

TokenBouncer request event tracing screen

Turn AI usage into operational policy.

Go beyond observing who uses which model. Enforce limits in the request path according to organizational structure and budget.

Single compatible control layer

Unify requests without changing how existing OpenAI-compatible clients and internal AI services connect.

Organization-aware enforcement

Combine global, group, user, and model scopes to inherit or override request, token, and budget limits.

Request-level operational records

Trace model, user, cost, latency, blocking, fallback, and request repair in a single event record.

Budget and model policies that match the organization

Reflect user and group inheritance, model-specific policy, and actual usage distribution instead of relying on one fixed limit.

Request, token, and budget limits

Limit monthly requests, tokens, and spend per user or across an entire group.

Model-specific policies

Set separate limits and priorities by model name or pattern across global, group, and user scopes.

Inheritance and overrides

Inherit default group policy while overriding limits or disabling access only where needed.

Usage-informed suggestions

Review policy candidates based on existing usage distribution and budget before an operator applies them.

Resolve the effective policy from priority and scope.

Lower priority numbers are evaluated first. Exact model names and longer patterns win ties, making the final result clear even when global defaults, group rules, and user exceptions overlap.

Global defaultP100 · all models · KRW 75,000 per user
Legal team policyP50 · claude-opus-* · KRW 3,000,000 per group
User overrideP10 · selected user · 300,000 tokens per month

TokenBouncer model fallback order settings

Track individual users and shared group budgets.

Connect groups from Microsoft Entra ID (Active Directory), Keycloak, and other internal identity systems to real request identities, then inspect inherited limits, individual overrides, and shared group budgets in one view. Regular users can also see their own utilization and reset time.

Per-user usage

Calculate confirmed usage, in-flight reservations, remaining limits, and utilization for requests, tokens, and spend.

Shared group budgets

Enforce department-wide shared budgets alongside personal limits to prevent organizational overspend.

Self-service usage view

Let non-admin users see their limits, utilization, and next reset time without exposing administration.

TokenBouncer individual user policy and usage details

We go beyond measurement and propose
cost optimization.

Unify provider pricing, actual input and output tokens, cache hits, and per-user usage patterns into one cost baseline. Compare model mixes by savings target, turn the selected scenario into limits and fallback policy, and keep validating realized savings after rollout.

Compare model budget mixes, projected total spend, and the share of users affected for each savings target, then turn the selected scenario into user, group, and model policies on the spot.

TokenBouncer model budget policy suggestions by savings target

Calculate the real cost baseline

Combine provider and model pricing with actual input, output, and cached tokens. Calculate confirmed model spend and cache savings on the same basis instead of relying on request counts.

Compare savings scenarios

Calculate premium, balanced, and efficient model budgets for savings targets from 10% to 90%. Compare projected total spend, spend per user, and the share of users affected.

Turn analysis into operating policy

Create user, group, and model budgets plus fallback order directly from the selected scenario. Keep premium models for critical work while moving repeatable tasks to efficient alternatives.

Verify savings after rollout

Continue comparing spend, model mix, and cache savings before and after policy rollout. Use latency and status events to confirm that savings do not create quality loss or excessive blocking. During a period, project the end-of-period cost from the current trend and the previous period average.

Cost allocation by task

Attach a task or project tag to a request to break down usage by tag, even for the same user. Tags need no advance registration; values are aggregated as they arrive, reporting spend by task or project.

Automatic prompt caching

Automatically add cache directives to Anthropic model requests to reuse repeated prompt prefixes. Operators can turn automatic caching on or off and check cache hit rate and savings on the dashboard.

See cost and failure causes
one operational flow.

Dashboard aggregates connect to request-level events, so cost changes and user support cases can be investigated from the same data.

Requests, tokens, cost, and cache savings

Combine model pricing with actual input and output tokens to calculate confirmed cost. Track tokens and spend saved through cache hits separately.

Analysis by provider, model, user, and group

Compare which teams and users rely on each model instead of seeing only total spend. Use concentration and budget burn rate to find where policy needs adjustment.

Latency, status, blocking, and fallback tracing

Record latency, status, and block reason for every request. When fallback runs, retain the requested and actual model, rule, reason, and attempt order.

Event CSV exports and administrator audit logs

Use an event ID to connect a support case to the exact request. Export events to CSV while retaining policy, key, and configuration changes in a separate admin audit log.

Operate through quota pressure and model failures.

Do more than block requests. Apply approved fallback models and retry policy, then record what happened.

Model fallback

When the original model reaches a limit, reevaluate approved alternatives in order within the same provider.

Retries and request repair

Retry upstream 5xx and transport failures, and safely repair known request-shape errors before one additional attempt.

Access keys and user identity

Combine virtual keys, trusted headers, and synchronization with Microsoft Entra ID (Active Directory), Keycloak, or another internal identity system to connect request credentials to real users.

MCP connection and audit

Manage the MCP servers and tokens available to AI agents and audit their connection and usage records.

TokenBouncer event table with model, tokens, cost, and latency per request

Issue auditable override tokens instead of temporarily raising limits.

Deadline work, incident response, and demos sometimes need to pass despite a limit. Raising the limit and lowering it back is easy to leave undone and records neither what changed nor why. A request carrying a policy override token skips only the quota decision; usage is still aggregated, and cost and event records are retained.

Expiry, audience, network, and use limits

Set expiration, allowed users and groups, network ranges (CIDR), and a use cap per token. Disabling or revoking a token takes effect immediately on every server.

No client changes required

Carry the token in a header, a body field, or a magic word typed into the message itself. Whichever path is used, the token is removed before the request goes upstream, so it never reaches external logs.

Bypasses capacity limits only

Only capacity limits are bypassed. A token cannot reactivate a disabled account or bypass Virtual Key validity checks and model allowlists.

Audit records for overrides

Events for overridden requests include an override badge, token details, and the limits excluded from evaluation. Failed overrides are also recorded with a reason, and revoking a token never erases past records.

Issue tokens with an expiry, allowed audience, network range, and use cap, then manage per-token override spend, limit inclusion, and last use in one place.

TokenBouncer policy override token management screen

Tune operational behavior to the environment.

Manage request handling, failure response, pricing data, and audit scope alongside policy. In restricted environments, disable remote pricing sync and use only the embedded snapshot.

Retries and request repair

Set upstream retry count and delay, and safely repair known request-shape errors.

Custom messages for each failure case

Customize messages for quota exhaustion, policy blocks, authentication failures, and upstream outages around organizational language and support procedure. Users receive a traceable event ID while operators retain the original error, retries, and fallback history.

Pricing and air-gapped operation

Use embedded model pricing or scheduled remote refresh. Keep cost calculation available without external sync in air-gapped networks.

Audit and export

Export request events to CSV and retain administrator changes, key activity, and MCP connections in audit logs.

Connect existing workspaces and identity systems.

Keep OpenAI-compatible APIs for internal AI applications and workspaces such as OpenWebUI. Combine OIDC, group synchronization, and trusted headers from Microsoft Entra ID (Active Directory), Keycloak, or another internal identity system so the existing account structure becomes the basis of AI policy. Deploy the proxy, administration interface, and database within your own infrastructure to preserve the current operating model.

Custom user header and access policy settings

Virtual Keys managed by purpose, model, and budget

OpenAI-compatible endpoint activation settings

Active endpoint URLs with a summary of the current configuration

Works with existing AI applications

Identify signed-in users from a custom user header and connect them to policy and events. Name the header to match the environment, or keep what an application already sends, such as OpenWebUI’s X-OpenWebUI-User-Email.

Internal identity systems

Import users, groups, and memberships from Microsoft Entra ID (Active Directory), Keycloak, or another internal identity system through OIDC and service-account sync, with admin access controlled by organizational groups.

Purpose-specific Virtual Keys

Separate keys for workspaces, batch jobs, and services, with allowed models, budgets, expiration, and last-use tracking.

Expose only required APIs

Enable selected OpenAI-compatible paths for chat, responses, embeddings, image, audio, and more by environment.

Store no sensitive content. Keep every operational record.

Request bodies are not stored by default; billing and quota decisions do not require prompt content. Usage events, audit logs, and original values saved before recalculation are retained without automatic deletion, preserving the records billing and audits need.

No prompt storage by default

Quota decisions and cost calculations use usage data only. An operator enables body capture explicitly, and only where verification requires it.

Horizontal scale and back-pressure

Add replicas to scale throughput. Connection pooling, intake limits, and back-pressure on the persistence queue prevent usage records from being dropped during surges.

Backups before recalculation

Before stored costs are recalculated with a new price table, the existing values are committed to an append-only backup. If the backup fails, recalculation does not start.

Operational changes without restarts

Policy, limit, and identity changes propagate to every server without a restart. Schema migrations run automatically at boot, so no maintenance window is needed.

All the control,
none of the slowdown.

Every request is inspected, yet the added latency stays around 0.2 ms — users cannot feel the difference.

+0.16 ms

Added latency per request (median)

Median round-trip difference between direct and proxied calls. Users under a budget limit see the same latency with default pass-first admission.

20,000+ RPS

Sustained throughput, single process

A single process served about 400k requests over 20 seconds at 50 connections with zero errors and a 4 ms p99 latency.

Zero loss

Usage records under overload

Beyond the intake cap, only excess requests receive 429. Not a single usage event from served requests was dropped.

Measured on a local Apple Silicon machine against an OpenAI-compatible local upstream. Absolute figures may vary by environment and hardware.

A major law firm already uses TokenBouncer firm-wide.

Yoon & Yang LLC has deployed TokenBouncer firm-wide, controlling AI usage and reducing cost across the entire organization. Lawnect works with partners to build AI operations that fit legal work.

Turn AI usage and cost into enforceable policy.

Tell us about your AI services, model providers, and organizational structure. We will propose the integration, policy scope, deployment, and operating model.

Control organizational AIthrough policy and cost.