Measures what is worth putting in context
You are paying for 600,000 tokens to get the value of 24,000
Teams pack the repository into the window and hope. Context Budget measures each chunk contribution to task success, prices it per token, and packs what earns its place.
The problem
Context is the biggest cost lever and the one chosen entirely by guesswork
Every agent framework offers a way to stuff more into the window and none of them tell you whether it helped. So teams include everything, pay for everything, and lose quality to the noise — because a lockfile in the context is not neutral, it is a distraction with a price tag.
The insight
This is a knapsack, and everyone is solving it by taking everything
Each chunk has a measurable contribution to task success and a known token price. That makes packing an optimisation with a correct answer, not a matter of taste. Greedy by value density gets within a hair of optimal here, and — unlike including everything — it can tell you exactly what it left out and why, which is the part that makes an engineer trust it.
Per-chunk contribution measured by ablation across a task set, value-density ordering under a token budget, and an explicit dead-weight set for chunks whose measured contribution is indistinguishable from zero.
How it works
Four steps, no data science team
Each chunk in and out across your task set, once. This is the only expensive step.
Contribution per thousand tokens, ranked.
You state the budget, it states what fits and what it dropped.
The repository moves and so does the ranking.
Who it is for
The engineer who owns the agent and the inference bill
Teams running coding or retrieval agents at volume, where context is the dominant cost and nobody has measured what it buys.
Pricing
- –Contribution measurement
- –Value-density ranking
- –Dead-weight report
- –Packing API
- –Continuous re-measurement
- –Budget policies
- –Spend reporting
- –Self-hosted
- –Per-workload budgets
- –SSO
- –Priority support
Competition
What exists, and what it does not do
| Who | What they do | The gap |
|---|---|---|
| Retrieval and reranking stacks | Select chunks by semantic similarity to the query. | Similarity is a proxy for usefulness and a poor one. Neither measures whether including a chunk actually changed the outcome. |
| Prompt caching | Cuts the price of tokens you send repeatedly. | Makes the wrong context cheaper rather than smaller. Complementary, not competing. |
| Agent frameworks with context management | Compaction, summarisation, sliding windows. | Heuristics applied uniformly. None of them know which of your files actually matters. |
| Send everything | The default, enabled by large windows. | Expensive, and the quality cost is invisible because nobody ran the comparison. |
The ablation is expensive to run and has to be re-run as the codebase moves, which puts a real cost in front of the value — a hard sell to someone who has not yet been surprised by their bill. The larger risk is structural: context windows keep growing and per-token prices keep falling, so the problem this solves is being eroded by the market. That argues for treating it as a feature of an agent platform rather than a company, and for selling it now rather than in two years.
Market
A share of inference spend, on the same budget line as Router
Real today at high volume and shrinking as windows grow. Best paired with Router, where one relationship covers both halves of the cost question.