Paritok

Paritok is a compression gateway for AI coding agents. You point the agent at a local proxy instead of the model provider — export ANTHROPIC_BASE_URL=http://127.0.0.1:8080 — and every request is rewritten on the way upstream so the provider bills for far fewer input tokens. Responses come back untouched, and the agent needs no modification.

The three levers it pulls

1. Tool-schema filtering. The largest single-turn win. With 46 tools registered, Paritok keeps the schemas plausibly relevant to the current turn in full and stubs the rest, reducing roughly 29K tokens of tool definitions to about 8K.
2. Content compression. File reads and tool output pass through a purpose-trained 4B model that keeps function signatures, imports and error strings while discarding filler, landing at about 26% of the original size.
3. History summarisation. Once the context budget fills, turns beyond a configurable recent window are summarised; recent turns stay verbatim.

Nothing is actually thrown away. Compression is "lossy on the wire, recoverable when it counts" — a read_original() call returns the exact bytes from local storage when the agent needs them.

The model behind it

Paritok-4B-v1 is a LoRA adapter over Qwen3-4B-Instruct-2507, trained on 45,000 real coding-agent trajectories so it can tell a function signature from a debug line. It needs roughly 2.5 GB at Q4 quantisation and runs on an 8 GB GPU, or CPU-only if you use tool filtering alone.

Reported results

On SWE-bench Lite the vendor reports 86.5% of uncompressed solve quality retained at a 25.7% compression rate, compared with 50.2% compression for GPT-4.1-mini prompt compression. Savings compound over a session — roughly 39% at five turns and 63% at twenty, per the project's own figures. These are the authors' benchmarks, not independent ones.

Compatibility and licensing

Works with Claude Code, Cursor, Codex, OpenHands and any OpenAI-compatible upstream. The gateway, the 4B model and the training scripts are all Apache-2.0; self-hosting is free and telemetry-free. A managed GPU endpoint is offered at $0.30 per million tokens for teams that would rather not run the model. Distributed via PyPI (pip install "paritok[proxy]"), Ollama and Hugging Face; current release is v1.3.0.