AX Optimize.

Measurably reduce LLM token costs — open-source-based context compression, plus cost governance.

DACH-wide

Agent runs balloon: tool outputs, logs, RAG chunks and long conversations flow into the context round after round — and every token costs money. AX Optimize reduces these costs measurably by compressing the context ballast before it reaches the model. In the real-world run above: 66.8% fewer tokens with answer quality preserved.

ax-optimize — real run

$ ax-optimize --file agent-session.json --price-per-mtok 3 --monthly-runs 5000

AX Optimize — Token Reduction
  Tokens:     16,049 → 5,321   (−66.8%)
  Saved:      10,728 tokens per run
  Transforms: log, smart_crusher  (user messages protected)

  Cost estimate (input tokens):
    at €3.00/1M tokens, 5,000 runs/month:
    ~€160.92/month  ·  ~€1,931.04/year

  # Reduction is input-dependent: tool outputs/logs high, dense prose low.

Honestly framed: the compression engine is Headroom (open source, Apache 2.0) — we did not invent it, we integrate and operate it. It protects your actual user messages and primarily compresses the machine-generated ballast. Our contribution (the AX delta) is the cost story: tokens saved translated into euros and projected onto your volume — plus model routing and cost caps for genuine budget governance. Local-first: the context does not leave your machine via us.

Context compression

Tool outputs, logs and RAG content are condensed via the Headroom engine — user messages remain protected.

Cost projection

Saved tokens are translated into euros and projected onto your runs per month — transparently labelled as an estimate.

Model routing & caps

Cheaper models for simple steps, hard cost caps per day/task — budget governance instead of a surprise on the invoice.

Local-first

Compression runs on your machine. The context does not flow out via us — relevant with sensitive data.

Measurable token reduction per run — typically 60–95% with tool-/log-heavy context
Euro projection onto your volume (month/year), transparently labelled as an estimate
Integration via MCP tool or as a proxy/library into your agent pipeline
Local-first: context stays on your infrastructure
Honest OSS basis: the engine is Headroom (Apache 2.0), attribution included — our value is integration, routing, cost governance

Does answer quality suffer from the compression?

Compression targets the ballast specifically — repeated tool outputs, logs, RAG chunks — not your actual prompts. User messages are protected. In practice, answer quality is preserved; the reduction depends heavily on the input.

How much does it really save?

For tool- and log-heavy agent runs, 60–95% of context tokens. For dense prose, less. The run on this page shows 66.8% on real agent context. We measure it against your data before we promise anything.

Does our context leave our own infrastructure?

No. Headroom is local-first; the compression runs on your machine. Nothing flows out via us.

What is your contribution compared to the open-source project?

The engine (Headroom, Apache 2.0) is OSS — we state that openly, including attribution. Our contribution is the integration, the model routing, the cost caps and the euro projection that turns compression into a budget decision.

Related roles & solutions

Goes together

Product AX Code-Review-Agent The live-demonstrable code review agent from the same suite. Overview Agent-Lösungen The entire AX Suite and bespoke solutions.

AX Optimize — does this fit your situation?

In a free initial call we clarify whether this service addresses your problem — or whether another route gets there faster. Honest assessment included.

Book an initial call