Agentic continuous optimization · always on

24x7 autonomous optimization of your AI workloads.

LensOp sits on top of your routers, your coding tools and your own models, and runs a continuous loop on every call: forecast, route, compact, cache, govern. No human in the loop. Cost per ticket, developer and feature comes along for free.

No signup · The demo runs on a real backlog, not screenshots

24x7Autonomous optimization, no human in the loop
every callForecast, routed and governed at the gate
30 secTo point your router at LensOp
0Code changes to get started
Live stream · one engineering org · 180 developers Move the lens →

Sample of a live call stream. Without LensOp each line shows only a date and an amount. With LensOp the same line resolves to the tool or model that made the call, the ticket it belongs to and the reason it cost what it did, for example a retry storm, a nightly CI agent with no cache reuse, duplicate context across sub-agents, an idle self-hosted GPU, or work re-run after rate limiting.

07 / 03claude-codeOC-1421 · retry storm, 6 hour agent session$4,180
07 / 14copilotOC-1508 · nightly CI agent, no cache reuse$7,412
07 / 22cursorOC-1547 · 3 sub-agents, duplicate context$9,034
07 / 28llama-70b · self-hostedOC-1560 · batch size 1, 41% idle GPU612 GPUh
07 / 31claude-codeOC-1588 · rate limited 14x, work re-run$12,844
What your dashboards show you today Attributed today 0% · with LensOp every call

LensOp works with Claude Code, Cursor, GitHub Copilot, Codex and Windsurf, with your own agents, with open weight models self-hosted on local GPU, and with LLM routers including LiteLLM, Portkey, Amazon Bedrock and Google Vertex AI.

The problem

AI runs 24x7. Your optimization doesn't.

Cloud had the same moment. Unbounded usage, optimized by hand, once in a while, until autonomous rightsizing took over and nobody went back. LensOp is that autonomous layer for AI workloads.

↳ Teams typically find their top 3 waste sources in the first week.

Routing decisions are made once, in code, and never revisited.
A single runaway agent loop can eat a month of budget overnight.
Optimization happens in quarterly cleanups, not on every call.
Nobody can prove whether last month's change made the output better or worse.
Your provider bill is one number. Everyone asks what it bought.
60%

of production LLM failures are rate limits. Flow breaks mid task and the work gets run again.

Datadog · 2026
19%

net slowdown on complex tasks with AI assistance, while developers believed they were 20% faster.

METR
98%

of FinOps practitioners now manage AI spend, up from 63% the year before.

FinOps Foundation · 2026
The blind spot

A proxy alone can't see the tools your team lives in.

Claude Code, Cursor and Copilot call the providers with their own keys. That traffic never reaches your router, so anything sitting only at the gateway is blind to it. It is not a missing feature, it is where the packets go. LensOp covers both paths, the traffic you route and the traffic you don't.

✕ Gateway only

Sees your API code. Misses most of the work.

Claude Codeinvisible
Cursorinvisible
Copilotinvisible
Self-hosted GPUinvisible
✓ With LensOp above it

One loop over every tool, agent, model and gateway you run.

Claude Coderead, attributed
Cursorread, attributed
Self-hosted GPUmeasured, batched
Your routerkept and steered

Running LiteLLM, Portkey or something homegrown? Keep it. LensOp is the decision layer above it.

The engine

One autonomous optimization loop, running around the clock.

Any AI workload in, any execution path out. LensOp observes every call, measures it, forecasts the next one, attributes it to a work item and governs it, then acts: route, compact, cache. Continuously, without anyone opening a dashboard.

Any AI workload
  • Coding agents
  • Chat and assistants
  • Autonomous agents
  • Data analysis
  • Search and retrieval
CONTINUOUS OPTIMIZATION LOOP OBSERVE MEASURE FORECAST ACT ATTRIBUTE GOVERN
Routeand fall back
Compactcontext
Cacheand reuse

Autonomous · always on, self-improving, adaptive · governed at the gate

Any execution path
  • Frontier APIs
  • Open weights, self-hosted
  • Local GPU, warm pool
  • Quantized tiers
  • Cache hit
The hard part

Anything can be made cheaper. Only a quality gate makes it safe.

Every change the loop proposes is scored against a baseline built from your own workloads. Cheaper only wins if it clears the bar. If it doesn't, the call falls back and the change never ships. That gate is why this can run on autopilot instead of sitting in someone's backlog.

Intent scoredcode-generation, high stakes
Candidatesmaller model, cached prefix
Quality checkclears baseline
Decisionshipped, logged
Next candidatebelow baseline, rolled back
What you get

Optimization first. Visibility as the by-product.

The loop does the work. The reporting below is what falls out of it, and it is all in the demo you can open right now.

24x7 autonomous optimization

The loop never stops. Understand intent, forecast the call, route to the cheapest path that clears the quality bar, compact context, reuse cache, and log every decision.

Self-governing guardrails

Quotas, priority lanes and rules act on their own. Alert, throttle or block before overspend, while hotfixes and critical agents keep headroom.

Anomalies caught and corrected

Context window blowups, retry storms, tool loops, oversized retrieval. Detected with the cause, then handled by the loop instead of written up in a postmortem.

Cost per work item

Every prompt is joined to a ticket, a PR and a story point estimate. See the true cost of a hotfix, a feature or a docs pass, planned against actual.

Developers as budget owners

Each engineer gets a daily allowance, an efficiency score per story point, and their own view of what they burned and on what.

Live stream with full traces

Watch calls land in real time as the team works, then click any tick to replay the full prompt trace and the decision behind it.

Before / after

Before LensOp, after LensOp.

Same team, same models. What changes is what you can see, cap and improve.

DimensionBeforeAfter
Optimization cadenceA manual cleanup once a quarterAutonomous, 24x7, every single call
RoutingModel chosen once, hardcoded, never revisitedIntent and stakes routed to the cheapest path that clears quality
QualityNobody knows if a change made output worseEvery change scored against a baseline before it ships
AttributionOne provider invoice, no ownerEvery call mapped to a ticket, developer and feature
Unit economicsNobody knows what a shipped feature costCost per work item and per story point, planned against actual
CoverageManaged tools and self-hosted models measured separately, or not at allOne loop across routed traffic, coding tools and your own GPUs
Budget controlOverspend discovered at month endTeam and category quotas with alert, throttle or block rules
Setup

Thirty seconds to first signal.

⌁ Option 1 · recommended

Spin up the proxy, let context flow

Point your router at the LensOp proxy. It continuously syncs tickets, new code, PRs and CI/CD signals into a live context graph, so attribution happens without a single tag in your code.

# 30 second setup
docker run -p 8787:8787 lensop/proxy
export OPENAI_BASE_URL=http://localhost:8787/v1
# connect Jira and GitHub in the console, done
>_ Option 2

Tag explicitly with the SDK

Prefer full control over attribution? Wrap your calls and pass intent, ticket and stakes yourself.

import { lensop } from "@lensop/sdk";

const res = await lensop.chat({
  intent: "code-generation",
  ticket: "OC-1421",
  stakes: "high",
  messages,
});

Claude Code, Cursor and Copilot don't need either. They never touch a proxy, so they connect read only from the admin side. Nothing to install on a developer's machine, nothing to change in a workflow. Self-hosted models connect through your existing metrics endpoint.

Design partner program

Put your AI optimization on autopilot this week.

Open the demo to watch the loop work a real backlog, or book a short call and we will map it to your own stack. Design partners get the first Workload Audit free, and the report is yours either way.

Good fit: 25+ developers on AI tools, two or more tools, or your own models in production