Context Compiler

WORKS IN YOUR AGENT · RUNS OFFLINE · NO KEY TO COMPILE

Same answer. About 3% of the tokens.

Give it a file, a question, and a token budget. It returns the useful parts as markdown, and a clear list of what it left out.

OpenAI × NamasteDev Codex Hackathon · July 2026 · solo build

One real handbook · one refunds question

Whole file

20,364 tokens

Compiled

591 tokens

97% fewer tokens

Same facts. Every read.

Checked with tests, a recall eval, and live runs

THE HABIT

Agents pay to read pages they don’t need

01

Whole-file habit

Agents often load an entire document to answer one question. Every time.

02

Most of it is unused

The useful clause might be 2 pages of 100. The other 98 are pure token spend.

03

It adds up fast

Every file × every question × every agent × every day. Cost and latency climb together.

$61 → $2

per 1,000 reads of one 100-page document (at $3 per million input tokens; demo cost meter default)

~95%

of file tokens often unused for a single task, across pdf, docx, xlsx, pptx, and images

FIT

Who it’s for, and where it fits

Plain version: who needs this, why we built it, what’s different, and how it sits next to converters.

Who

People building coding and document agents. Cursor, Claude Code, Codex, and internal agent platforms.

Why it exists

Whole-file reads are getting expensive. Agents keep reopening the same manuals and reports for different questions.

What’s different

One job. Pick what the agent needs under a token budget. Not a full search platform.

Where it fits

Converters turn a file into text of the whole thing. We select under a budget and tell you what was left out, so you can pull more if needed.

PIPELINE

File in. Useful parts out. Under your budget.

You set the ceiling. The agent gets markdown it can read, plus a list of skipped sections.

1

Convert

Turn the file into markdown. Cache by content so repeats are cheap.

2

Split

Break on headings. Keep tables whole.

3

Rank

Match sections to the question on your machine. No model call to compile.

4

Pack

Fill the strongest sections first until the question is covered, then stop.

Plug in, or try it here

Works with Cursor, Claude Code, Codex, and Claude Desktop. Or use the hosted web demo. Compile needs no API key.

Honest about cuts

Every result lists what was left out. Open a skipped section when you need more of a chapter. Nothing is trimmed in silence.

DEMO EVIDENCE

What you actually get back

Real 78-section handbook. One refunds question. Budget 1,200 tokens.

97%

fewer tokens, answer still intact

20,364 → 591

~34×

cheaper per read when conversion is cached; repeats come back fast

Check it

Same question from the full file and from the compiled slice. Compare the facts yourself.

Story: vendor handbook · “How long do refunds take and who approves large ones?” · budget 1,200

Whole file: 20,364 tokens

Compiled: 591 tokens · answer intact · list of skipped sections included

ROADMAP

What’s next

01

Clearer paraphrase matching

Local embeddings next. Query cleanup and offline recall eval already shipped.

02

Whole folders, not one file

Compile across a folder. Shared conversion cache for teams.

03

Video and audio

Transcribe into the same pipeline.

04

One-command install

Quick setup for any agent client that speaks the same protocol.

Try it live

context-compiler.onrender.com

​

github.com/shenba1712/context-compiler

Open the demo, or wire it into Cursor, Claude Code, or Codex. Check the answer on a real doc.

Stop paying for pages your agent doesn’t read.