ContextCrumb

LLM context gets messy fast: meeting notes, logs, issue threads, docs, transcripts, and tool descriptions all pile up until the useful signal is buried under filler.

ContextCrumb is a token-level compressor for LLM and agent workflows. It looks at text word by word and removes low-signal tokens while keeping the surviving text in the original order.

That is the idea behind the name: the context is still there, but the loose crumbs are shaken off before they reach your model. Less bloat in the prompt. More room for the parts that matter.

ContextCrumb is not a summarizer. It does not rewrite your document into a new explanation. It keeps the source sequence and deletes expendable words.

Original

ContextCrumb is designed for coding agents, MCP tools, and prompt pipelines that need to read a large local text file before sending it to an LLM. It prints only the compressed text by default, so an agent can capture stdout and use it as shortened context.

Compressed

ContextCrumb designed coding agents, MCP tools, prompt pipelines need read large local text file before sending LLM. Prints compressed text default, agent capture stdout use shortened context.

Same order. Less padding. More room for the next file.

Why ContextCrumb?

Best fit: docs, notes, transcripts, issue threads, logs, research context, and other natural-language files. For source code where exact syntax matters, prefer raw file loading or use a conservative keep ratio.

pip install contextcrumb

Optional extras:

pip install "contextcrumb[mcp]"
pip install "contextcrumb[serve]"
pip install "contextcrumb[torch]"

ContextCrumb uses the ONNX backend by default, so normal users do not need PyTorch or Transformers installed. Model files are cached locally after the first download.

from contextcrumb import ContextCompressor

compressor = ContextCompressor()

result = compressor.compress(
    "ContextCrumb deletes low-value words while preserving useful context.",
)

print(result.text)
print(result.stats)

Read and compress a file:

from contextcrumb import ContextCompressor

compressor = ContextCompressor()
result = compressor.compress_file("notes.txt")

print(result.text)
print(result.stats["token_keep_ratio"])

The main agent-friendly command is load:

contextcrumb load notes.txt

It prints only compressed text by default, which makes it easy for agents, hooks, shell scripts, and prompt pipelines to capture stdout and move on.

Useful commands:

contextcrumb load notes.txt --json
contextcrumb diff notes.txt
contextcrumb inspect notes.txt
contextcrumb stats

diff marks deleted tokens like this:

kept words [-deleted words-] kept words

ContextCrumb includes an optional MCP stdio adapter for agent clients that can run Python tools through uvx.

pip install "contextcrumb[mcp]"

Published-package MCP config:

{
  "mcpServers": {
    "contextcrumb": {
      "command": "uvx",
      "args": [
        "--from",
        "contextcrumb[mcp]",
        "contextcrumb-mcp"
      ]
    }
  }
}

The MCP server exposes:

compress_text
compress_file

ContextCrumb also ships contextcrumb-shrink, an MCP proxy that compresses verbose catalog descriptions before an agent sees them while forwarding tool names, schemas, calls, results, and resource contents unchanged.

Model weights and a hosted demo are public on Hugging Face:

• Model: ymao20/contextcrumb-32m

• Playground: contextcrumb-32m-demo

Roadmap

Planned for later:

• Public docs for advanced compression modes and service deployment.

• JavaScript or TypeScript client.

• Hosted API experiments.

• npm publishing.

Development

uv pip install --python .\.venv\Scripts\python.exe -e ".[dev,mcp]"
.\.venv\Scripts\python.exe -m pytest
.\.venv\Scripts\python.exe -m build

Release notes are tracked in CHANGELOG.md.

License

Apache-2.0. See LICENSE.