Why the limit runs out
- The model re-reads the whole conversation for every question.
- Tool results, logs and JSON output stay inside that conversation.
- The agent starts learning the codebase from scratch in every new session.
- Code you never asked for gets written, and then that code enters the context too.
So the limit is eaten from four different places. Each tool below closes one of them. That is why the reel lists all four together.
01 · Headroom
Compresses before it reaches Claude
Sits between you and the model and compresses unnecessary content before it ever reaches Claude. Works as a library, a proxy or an MCP server.
github.com/headroomlabs-ai/headroom (Open on GitHub)
Install
uv tool install --python 3.13 "headroom-ai[all]"
pip install "headroom-ai[all]"Either command is enough. The uv route builds an isolated environment for the tool, so your system Python stays untouched.
The number
Its README promises 60-95% fewer tokens on JSON data and 15-20% fewer on coding agents. That is the tool's own claim, we did not measure it.
When you need it
When tool results, logs and JSON responses are long. That is, when half the conversation is machine output.
Careful. GitHub hosts several projects with the same name. The compression one is the address above; install another and you gain nothing.
02 · Graphify
Reads the code once, not every time
Turns a codebase into a knowledge graph so Claude does not re-read the code every session. The graph lives in a file and survives the session.
github.com/Graphify-Labs/graphify (Open on GitHub)
Install
pip install graphifyy && graphify installThe package is named graphifyy, with two y letters. The command is still graphify. On macOS use pipx install graphifyy if pip fails.
The number
Its README states 71.5x fewer tokens per query compared to reading the raw files. That is the tool's own claim. Our own measurement is in the section below.
When you need it
When the project is large and the same questions keep coming back.
03 · CodeBurn
A bank statement for your spend
Shows where your AI coding spend goes. Think of it as a bank statement: where the money went, which model, which project.
github.com/getagentseal/codeburn (Open on GitHub)
Install
npx codeburn
npm install -g codeburnThe npx option runs it once without installing anything. Use the second command for permanent use.
The number
We publish no savings percentage here, because this tool measures, it does not save. Its value is that only it can show what the other three actually did.
When you need it
Install this one first. Without knowing what you spend, you cannot know what the other three gave you.
04 · Ponytail
Code never written burns no tokens
Pushes the agent to write less code. The cheapest token is the one never written, so this tool attacks the problem from a different side than the other three.
github.com/DietrichGebert/ponytail (Open on GitHub)
Install
/plugin marketplace add DietrichGebert/ponytailThis command is typed inside Claude Code, not in a terminal. The repo lists separate lines for Codex and Copilot.
The number
Its README states roughly 54% less code, 20% cheaper and 27% faster, measured across 12 tasks with Haiku 4.5, n=4. That is the tool's own benchmark, we did not reproduce it.
When you need it
When the agent starts building things you never asked for.
Careful. The reel ran out of time before this tool, so everything here comes from the repo README, not from our own use.
Measurement
We measured this on our own codebase
We built a graph of our own Instagram project: 164 files, roughly 568 thousand words.
The one-off cost was 1,248,812 input tokens. 15 agents ran, averaging 83,254 tokens each.
The result: 2,222 nodes, 4,862 edges, 125 communities. 92% of the extraction came straight from the text, 7% was inferred.
What cost what
| What | Tokens |
|---|---|
| Total, 15 agents | 1,248,812 |
| Average per agent | 83,254 |
| 1 image, in its own agent | 74,979 |
| 8 images, in one agent | 79,685 |
| 6 boilerplate SVGs, result 6 nodes | 67,704 |
| 15 documents, result 107 nodes | 186,296 |
An honest note: 1,248,812 is the real sum across all 15 agents. A figure of 1,102,516 was visible for a while; it was a hand-typed mistake made during the merge step and has been corrected.
Three things that did not work
- A single vendored library file took up 32% of the graph, 721 nodes. None of them meant anything.
- 6 boilerplate SVG files cost 67,704 tokens and produced 6 nodes in total.
- A single image in its own agent yields 15-26 nodes. Batch 8 images into one agent and you get 1 node per image. Cheap, but empty.
The rule: the price is set by the number of agents, not by file size.
When none of this helps
- If you do not write code, none of these tools are for you.
- On a small project the graph never pays for itself: the one-off cost outgrows the questions you will ask.
- The graph is static. It goes stale when the code changes and has to be rebuilt.
- If your limit comes from an hourly cap or a plan tier rather than tokens, compression will not lift it.
- Three of the four numbers are the tools' own claims. We only measured Graphify on our own codebase.
If you do not write code
Whoever hides the number is ashamed of it
If you read this page and decided it is not for you, one thing remains.
We published the 1.2 million tokens our own graph cost: what it earned, and what it turned into noise.
The team that builds your website should report the same way: how many pages, how many days, what did not work.
Do you know which page of your site actually brings you customers?
