Stop AI overload before it becomesan enterprise incident.
MoFlux helps platform teams coordinate shared LLM capacity, reject unsafe work before it reaches providers, and explain every admission decision. Today, Latchflo distributes expiring capacity grants and Tyr enforces concurrency and in-flight token admission limits locally—without putting the control plane on every request.
Current protocolSQLite state plus HTTP/JSON polling
Enterprise pain points
AI capacity failures rarely stay isolated.
LLM requests vary dramatically in size and duration. When capacity is managed independently by each replica—or hidden behind growing queues—the result can be degraded latency, provider pressure, retry amplification, and incidents that are difficult to explain.
Available today
Traffic spikes become cascading failures
Business impact
Large or concurrent LLM calls can exhaust downstream capacity, increase tail latency, and spread instability across dependent services.
How MoFlux responds
Tyr evaluates concurrency and estimated in-flight token demand before the provider call begins, then admits, queues within a configured bound, observes, or rejects the work.
Available today
Replica-local limits do not protect a shared fleet
Business impact
A safe limit on one gateway can become unsafe when dozens of replicas independently apply the same ceiling.
How MoFlux responds
Latchflo partitions a fleet-wide capacity pool into bounded, expiring per-agent grants, while Tyr enforces only the capacity assigned to its local process.
Available today
Queues hide overload until latency is already unacceptable
Business impact
Unbounded waiting turns capacity pressure into timeouts, retry storms, memory growth, and unpredictable user experience.
How MoFlux responds
MoFlux is admission-first. Tyr can fail fast by default or permit only a deliberately bounded queue, keeping overload visible instead of silently accumulating work.
Available today
Runtime policy changes drift or partially apply
Business impact
Changing concurrency, queue, and token settings independently can leave replicas in inconsistent or unsafe intermediate states.
How MoFlux responds
Tyr applies complete, versioned limit snapshots atomically and rejects stale revisions. Latchflo records controller epochs, grant revisions, acknowledgements, and lease expiration.
Available today
Teams cannot explain why a request was rejected
Business impact
Without decision context, platform teams struggle to tune policies, respond to incidents, or prove which configuration governed a request.
How MoFlux responds
Tyr exposes admission mode, decision reason, limit revision, admission identity, and—when managed by Latchflo—the grant ID and controller epoch tied to the decision.
Available today
A centralized controller can become a new request-path dependency
Business impact
If every provider call requires a remote policy lookup, control-plane latency or downtime can directly affect application availability.
How MoFlux responds
Latchflo remains off the live request path. Tyr uses locally installed grants and continues making admission decisions until that authority expires, then fails closed.
What changes with MoFlux
Capacity becomes an explicit operating policy.
Instead of allowing every replica to discover overload independently, MoFlux separates fleet coordination from local execution and makes the resulting decisions visible.
Protect service health
Reject or observe work before it consumes scarce provider-facing concurrency and in-flight token capacity.
Coordinate shared capacity
Partition global limits across active Tyr agents instead of multiplying the same ceiling across every replica.
Roll out with evidence
Use observe mode to measure prospective rejections before enabling enforcement for production traffic.
Fit the existing platform
Route only selected Anthropic or OpenAI traffic through Tyr while retaining existing ingress, auth, networking, and secrets.
Enterprise outcomes
Different teams gain a shared control model.
MoFlux is designed to make capacity policy operationally useful across engineering, reliability, product, and governance functions.
Platform engineering
One capacity authority for a fleet of local gateways without a per-request control-plane round trip.
SRE and operations
Explicit rejection reasons, bounded admission behavior, health endpoints, live statistics, and grant-level provenance.
AI product teams
A measured path from observe mode to enforcement without changing provider request formats or surrendering credential ownership.
Security and governance
Versioned policy application and an auditable control trail today, with stronger enterprise identity and access controls on the roadmap.
The MoFlux system
Central authority. Local enforcement.
Latchflo and Tyr solve different parts of the same enterprise problem: coordinate shared capacity centrally, but keep request-path admission fast and local.
Prototype available
Control plane
Latchflo
Latchflo owns the fleet-wide capacity view. It registers Tyr agents, allocates equal shares from configured pools, issues versioned expiring grants, receives heartbeats and acknowledgements, and preserves a durable audit trail.
Global concurrency and optional in-flight token pools
Unallocated safety reserve and high-priority token reserve
Durable controller epochs and odd/even revision safety model
Lease-aware redistribution that avoids early capacity reuse
Administrative HTTP API, health probes, stats, events, and Prometheus metrics
Current boundary: single controller, SQLite, bearer-token administration, and HTTP polling—not yet a production HA enterprise control plane.
Runtime available
Data plane
Tyr
Tyr is the admission-first runtime gateway. It estimates request demand, evaluates the local versioned limit snapshot, and either proxies, observes, queues within a bound, or rejects the provider call before scarce capacity is consumed.
Concurrency and in-flight token-aware admission
Fail-fast enforcement or shadow observe mode
Anthropic Messages and OpenAI Chat Completions proxy routes
Streaming usage reconciliation and adaptive input estimation
Atomic limit updates, stable admission IDs, and grant provenance
Current boundary: single-process gateway instances, two provider-shaped APIs, no provider-format translation, and no billing-grade cost enforcement.
Runtime foundation
async-bulkhead-llm 3.11.1
Tyr uses a token-aware admission engine with immutable reservation previews, admission-linearized revisions, observe mode, adaptive estimation, stable admission IDs, usage reporting, and bounded drain behavior.
Enterprise relevanceThe same local runtime that admits work also accounts for its capacity and applies control-plane revisions.
Current-state architecture
The working control loop today.
The control path distributes authority; the request path consumes it locally. Latchflo does not participate in every LLM call, and an expired grant causes the local agent to stop new admissions.
Current implementation: Latchflo coordinates versioned, expiring capacity grants while Tyr performs local token-aware admission and proxies supported provider requests.
Durable authority
Controller epochs, pool revisions, agent records, grants, ACKs, and audit events survive process restart in SQLite.
Local safety boundary
Tyr applies complete snapshots atomically and fails closed at lease expiration rather than operating indefinitely on stale authority.
No remote lookup per call
Admission remains a local runtime decision, avoiding a new synchronous control-plane dependency in the provider request path.
Integration architecture
Adopt MoFlux around the systems already in place.
A practical enterprise deployment keeps existing ingress, application auth, service networking, provider credentials, and observability. Only selected outbound LLM traffic is routed through Tyr, while Latchflo coordinates the gateway fleet outside the request path.
Example integration: insert Tyr in front of selected LLM-bound traffic, retain existing platform controls, and use Latchflo to coordinate capacity across the Tyr fleet.
Low-risk enterprise adoption
Start with evidence, not a hard cutover.
Tyr's observe mode runs the same reservation and capacity decision while allowing would-be capacity rejections to continue upstream. Teams can quantify policy impact before enforcement and expand scope incrementally.
Observe mode is not a load simulator. It measures the policy that would exist under enforcement while bypassed requests continue to create real upstream load.
01
Select a bounded workload
Start with one LLM-backed service, provider route, model group, or non-critical environment rather than changing the entire estate.
02
Insert Tyr into the outbound path
Keep provider-shaped request bodies and existing credentials. Route Anthropic Messages or OpenAI Chat Completions through Tyr.
03
Measure in observe mode
Run the real reservation and admission logic without blocking capacity-related requests, then tune limits against actual workload behavior.
04
Enable enforcement
Turn on fail-fast or bounded-queue admission after the policy is validated, while monitoring rejection reasons and pool pressure.
05
Coordinate the fleet with Latchflo
Register Tyr agents and distribute versioned, expiring grants from a central capacity pool outside the request path.
Current enterprise boundaries
What MoFlux should not be represented as—yet.
The present system demonstrates the distributed capacity model, but several controls expected in a mature enterprise platform remain roadmap work. The website distinguishes those boundaries directly.
Not yet a production HA control plane
Latchflo currently runs as a single-controller prototype with SQLite. Multi-controller coordination and PostgreSQL are roadmap work.
Not a billing system
Today's token budgets limit concurrent in-flight exposure. They are not monthly spend caps, provider invoices, or billing-grade token accounting.
Not a universal AI gateway
Tyr currently supports Anthropic Messages and OpenAI Chat Completions shapes and does not translate between provider formats.
Enterprise identity is still ahead
The prototype uses bearer-token administration and agent credentials. OIDC, RBAC, workload-attested mTLS, and policy governance are future capabilities.
Present foundation, enterprise roadmap
Build from a working safety model toward a governed platform.
The current column reflects the uploaded implementations. Near-term and longer-term columns describe intended production evolution, not shipped capability.
Available now
Working capacity foundation
Latchflo 0.3.0 single-controller prototype
SQLite-backed epochs, pools, agents, grants, acknowledgements, and audit events
HTTP/JSON registration, heartbeat, desired-state polling, and grant ACK
Equal-share concurrency allocation with optional in-flight token budgets and safety reserve
Expiring leases, safe capacity reuse, timeout reconciliation, and local fail-closed behavior
Prometheus-format Latchflo metrics and administrative inspection endpoints
Anthropic Messages and OpenAI Chat Completions proxy routes
Enforce and shadow observe modes, bounded queueing, and priority token reserve
Streaming usage reconciliation, adaptive estimation, live stats, and provenance headers
Near-term direction
Enterprise hardening
PostgreSQL-backed control-plane state
Workload identity, mTLS, OIDC, and role-based administration
OpenTelemetry exporters and richer metrics, logs, and traces
Canary rollout, rollback, and acknowledged two-phase grant handoff
Deployment packaging for common enterprise platforms
Operational dashboards, alerts, and policy-management workflows
Longer-term direction
MoFlux platform
Highly available multi-controller operation
Streaming or push-based control updates
Regional capacity envelopes and placement-aware allocation
Provider RPM, TPM, retry, and billing-oriented cost budgets
Broader upstream protocols and workload classes
Managed MoFlux control-plane and governance experience
Make overload a controlled decision—not an accidental outage.
MoFlux is built around a clear enterprise operating model: define fleet capacity centrally, enforce admission locally, preserve request-path independence, and make every policy revision and decision attributable.