⌁
Tool calls dump everything.
Search results, logs, and intermediate output often travel into subsequent turns whether they're still useful or not.
AI EXECUTION OPTIMIZATION
Stop paying your AI to reread everything.
Old conversations. Repeated tool results. Documents the model doesn't need. MinToks is building the layer that decides what context deserves to go, which model should handle it, and whether the result is worth the cost.
Available now: free, inspectable text optimizer & local CLI / Broader platform in development
THE CONTEXT PROBLEMILLUSTRATIVE FLOW · PRODUCT VISION
AGENT CONTEXTWHAT GETS SENT?
InstructionsPRESERVE
Repeated tool logsREVIEW
Relevant evidenceKEEP
Stale historyREVIEW
Reusable cache prefixPROTECT
↓
m/t MINTOKSPOLICY · QUALITY · COST
↓
THE MODELThe context it actually needs.
LESS REPEATED CONTEXT✳BETTER COST PER TASK✳QUALITY BEFORE RATIOS
THE COST IS IN THE CONTEXT
Your agent is doing the job.
Its context is doing overtime.
Adding more context to every turn is easy. Paying for it, managing its relevance, and making sure it doesn't drown out the important parts is harder.
TOOL OUTPUT
⟳
History accumulates.
Long-running agents need decisions and facts, not necessarily the entire transcript that produced them.
◇
Cheap tokens aren't always cheaper.
Compression can destroy cache reuse. A shorter prompt can cost more. The bill, latency, and task result all matter.
THE PRODUCT WE'RE BUILDING
A smarter decision
before every model call.
Decide what context to send, which model and provider should handle it, and when a change really improves the outcome. Sometimes the best decision is to leave it alone.
YOUR APPLICATIONAgent / RAG / WorkflowAlready running
MINTOKSContext policyInspect · Decide · Check
YOUR CHOICEModel / ProviderDirect, gateway, or private
IN RESEARCH
Context selection
Find and retain relevant evidence instead of passing along every retrieved paragraph or tool result.
IN RESEARCH
Memory efficiency
Use the facts an agent needs now. Keep provenance and changing decisions in view.
PROTOTYPE
Model & provider routing
Evaluate models and providers against cost, latency, permitted regions, capabilities, and session continuity. The current selector runs offline only.
PLANNED
Cache-aware choices
Account for cached input, stable prefixes, and provider pricing before declaring a win.
PLANNED
Outcome checks
Compare cost and latency alongside answer quality. If a strategy makes the task worse, it shouldn't ship.
PLANNED
Portable execution
Use direct APIs, an existing gateway, or private and local inference. The long-term policy should be yours to run and move.
Works with your stack, by design.Routing could work with your existing gateway or eventually execute through MinToks directly. Neither path should lock your policies to a vendor.
WORKING PREVIEW · NO MODEL CALLS
Try it on actual text.
This is a real, small piece of the stack: deterministic cleanup of repeat content, progress output, and JSON whitespace. Inspect every change. No invented AI compression magic.
ORIGINAL SIZE0characters
TEXT REMOVED0characters
RELATIVE REDUCTION0%of characters
What this preview proves
These transformations run and return inspectable text. They do not establish improved model performance, token-billing savings, or enterprise-grade safety.
What it doesn't
The endpoint does not use a model-specific tokenizer. The four-character token estimate returned by the API is only an approximation. Review changes before reuse.
THE METRIC THAT MATTERS
Not the shortest prompt.
The cheapest successful task.
Tokens removed are a useful diagnostic. They aren't a business outcome. Real optimization has to account for cache pricing, extra processing, latency, and whether the model still gets the answer right.
↘
The MinToks thesis is to optimize the whole call, not chase a bigger compression percentage.
EXPLORE THE ECONOMICSILLUSTRATIVE ONLY
Enter a hypothetical workload. This is a cost scenario, not a tested MinToks savings claim.
MODELED MONTHLY INPUT SPEND AVOIDED*$0Calculating scenario…
*30 days. Assumes a 10% cache-read rate, unchanged cached prefix, and that reduction applies only to uncached input. Excludes cache writes, processing overhead, output tokens, latency, and accuracy changes. Provider economics vary. No MinToks performance estimate is implied.
GOOD QUESTIONS
Before you put it in your stack.
These are the things we'd ask before trusting another component with production AI traffic.
Often, yes. Stable prefixes can be much cheaper on a cache hit. But caching doesn't decide which search results, logs, or historical facts deserve to enter the next request. MinToks is designed to evaluate both, because a shorter prompt can cost more if it breaks cache reuse.
They already solve important parts of the problem. MinToks aims to coordinate context, model, and provider choices using an existing gateway, direct endpoints, or an optional future MinToks route. The policy should stay independent of the transport vendor.
The only defensible way is to test. Our roadmap calls for baseline comparisons, source provenance, task-success checks, and a safe no-change fallback. The current preview is a deterministic cleanup utility, not a verified quality-preserving optimization gateway.
Not yet. The public demo sends text to a Cloudflare Worker, processes it in memory, and does not write it to a MinToks application database. It has no tenant authentication or enterprise security certification. Please don't paste confidential data or credentials.
FOUNDING PILOTS
Bring your most expensive agent loop.
We're looking for real workloads where agent context is getting out of hand. Let's establish the baseline, identify what can safely change, and see whether the economics improve.
Discuss a founding pilot No performance promises before the measurements.