tokenfold 0.1.0

Token-compression toolkit for LLM payloads (Python binding over the tokenfold-core Rust crate)

tokenfold

Fold the noise. Keep the signal.

The local compression layer for LLM apps and agents. Shrink messages, tool schemas, logs, and diffs before inference—with exact token counts and a safety receipt for every change.

Quick start · Why tokenfold · Integrations · Safety · Benchmark

45.63% fewer tokens on the checked-in structured-JSON benchmark, measured with the exact tokenizer. Reproduce it with cargo bench -p tokenfold-core.

LLM systems routinely spend context on formatting, repeated logs, verbose schemas, and low-signal command output. That waste costs money and crowds useful information out of the context window.

tokenfold sits between your application and any model provider:

app or agent  ──  messages · schemas · logs · diffs  ──▶  tokenfold  ──▶  any LLM
                                                            │
                                                            └── payload + receipt

It removes what the model does not need, protects secret-shaped values first, and reports exactly what changed. Preview every transformation before applying it.

Quick start

tokenfold is currently source-only. Build the CLI with the pinned Rust toolchain:

git clone https://github.com/snchimata/tokenfold.git
cd tokenfold
cargo build --release --locked -p tokenfold-cli

Preview savings on the bundled OpenAI payload without changing it:

target/release/tokenfold inspect examples/openai_payload.json --format openai

The abridged receipt:

TRANSFORM              MODE          EST TOKENS BEFORE→AFTER      SAVED       %  STATUS
secret_redaction       all                           346→346          0    0.0%  no_op
json_minify            all                           346→229        117   33.8%  applied
schema_compaction      all                           229→213         16    7.0%  applied
TOTAL  346 → 213 tokens   saved 133 (38.4% reduction)

Compress to a budget. The payload goes to stdout and the receipt goes to stderr, so the result is safe to pipe:

target/release/tokenfold compress examples/openai_payload.json \
  --format openai \
  --target-tokens 250 > compressed.json

Or trim noisy output from any command:

target/release/tokenfold wrap -- git diff

Why tokenfold

• Spend context on the task. Remove formatting, repetition, and low-signal content before it consumes the model's attention or your budget.
• Know what you saved. Exact o200k_base and cl100k_base counts drive budget decisions; heuristics are never presented as exact.
• Keep control. Every call returns a typed CompressionReport with before/after counts, per-transform savings, estimator provenance, warnings, and final status.
• Fail honestly. If the safe transform set cannot meet a target, tokenfold reports BEST_EFFORT or UNREACHABLE_TARGET instead of pretending it succeeded.
• Stay provider-neutral. Use the same Rust core from the CLI, Python, an HTTP proxy, an MCP server, or your own Rust application.
• Run locally. The CLI is a native binary with no service account, hosted control plane, or runtime dependency.

One engine, five ways to use it

Surface	Entry point	Best for
CLI	tokenfold	Files, stdin, command output, and local agent workflows
Python	import tokenfold	Python applications and evaluation pipelines
HTTP proxy	tokenfold-proxy	OpenAI- or Anthropic-shaped traffic in flight
MCP server	tokenfold mcp serve	MCP-compatible agents and editors
Rust library	tokenfold-core	Native embedding with typed policies and reports

Python

The Python package is a native pyo3 binding for CPython 3.9+. Until wheels are published, install it locally with maturin:

python -m venv .venv
source .venv/bin/activate  # PowerShell: .venv\Scripts\Activate.ps1
python -m pip install maturin
maturin develop --release -m crates/tokenfold-py/Cargo.toml

from pathlib import Path
from tokenfold import CompressionMode, compress_openai_payload

payload = Path("examples/openai_payload.json").read_text()
result = compress_openai_payload(
    payload,
    target_tokens=250,
    mode=CompressionMode.BALANCED,
)

print(result.payload.decode())
print(f"saved {result.report.saved_tokens} tokens ({result.saved_pct():.1f}%)")

Also available: compress, inspect, compress_anthropic_payload, and compress_messages. Errors use a typed hierarchy rooted at TokenFoldError.

HTTP proxy

cargo build --release --locked -p tokenfold-proxy
target/release/tokenfold-proxy \
  --upstream https://api.openai.com \
  --target-tokens 12000

The proxy binds to 127.0.0.1:8787 by default, requires an HTTPS upstream, streams SSE without buffering, and returns X-TokenFold-* receipt headers.

MCP server

target/release/tokenfold mcp serve

The stdio server exposes tokenfold_compress, tokenfold_inspect, tokenfold_retrieve, and tokenfold_stats over JSON-RPC 2.0.

Safe by design

tokenfold applies ordered transforms, recounts after every stage, and stops when the target is met or the safe transform set is exhausted.

input  ──▶  redact secrets  ──▶  compact  ──▶  recount  ──▶  payload + report

Stage	Availability	What it does
secret_redaction	Always on	Redacts secret-shaped values before reporting, logging, or persistence
json_minify	Default	Removes insignificant JSON whitespace
schema_compaction	Default	Shortens examples while preserving descriptions and required fields
log_compaction	Balanced / aggressive	Collapses repeated, low-signal log lines
diff_compaction	Experimental	Reduces low-signal diff context

Lossy transforms must clear a downstream fidelity gate before promotion. Originals can be stored by SHA-256 hash when reversibility is requested; secret-shaped content is never persisted.

Measured, not marketed

The structured-JSON benchmark measures 45.63% exact token savings in balanced mode on a deterministic, pretty-printed OpenAI tool-schema fixture. It is a reproducible benchmark, not a promise for every workload.

cargo bench -p tokenfold-core

The tokenizer, fixture, hardware record, and regression thresholds are documented in crates/tokenfold-core/benches/THRESHOLDS.toml and CHANGELOG.md.

CLI command map

Command	Purpose
inspect	Preview achievable savings
compress	Compress a file or stdin
diff	Compare raw and compressed payloads
wrap (shell)	Run a command and compress its output
benchmark	Measure fixtures with before/after token counts
init / uninit	Install or remove an agent-host integration
doctor	Check estimator, config, and integration health
mcp serve	Start the MCP server over stdio
retrieve	Restore an original saved by hash
stats / gain	Report local token savings
session	Report command-wrapping coverage
filters	Manage command-output filters
output-savings	Report measured or estimated output-token savings
learn (discover)	Propose policy changes from local history

Run tokenfold <command> --help for flags and examples.

Status

tokenfold is source-only 0.1.0 software. Prebuilt binaries, crates.io packages, and PyPI wheels have not been published yet. The CLI, Python binding, proxy, MCP server, and optional extension crates are implemented and tested in this repository.

Contributing

Issues and pull requests are welcome. Run the same core checks as CI before opening a PR:

cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace --locked
python eval/run_fidelity.py --gate --profile smoke-first-consumer

Security and dependency checks:

cargo audit
cargo deny check advisories bans licenses sources

License

Apache-2.0