Paritok
A context gateway that provides real-time, non-destructive compression of tools, files, and history for AI coding agents
About Paritok
Paritok is a compression gateway for AI coding agents that performs real-time, non-destructive compression of tools, files, and history; the official claim is longer sessions and smaller bills. Point ANTHROPIC_BASE_URL to a local gateway to connect, with support for Claude Code, Cursor, Codex, OpenHands, and OpenAI-compatible upstreams. For the 70+ tools in each request, it keeps relevant items and stubs the rest; a 4B model compresses file reads and outputs while preserving identifiers, paths, and error messages, and original bytes can be retrieved locally. The entire stack is Apache-2.0 open source and self-hostable, with tool filtering running on CPU; a managed endpoint is also available at $0.30 per million tokens, free before October 1.
Best for
Best use cases
Compress tools, files, and history during long coding sessions to extend sessions and reduce bills
When agent requests carry many tool definitions, keep relevant tools and stub the rest to reduce request size
Compress file reads and output transfers while preserving identifiers, paths, and error messages
Retrieve exact original bytes locally on demand without consuming extra turns
Features
Real-time non-destructive compression
Compresses tools, files, and history during requests; the official claim is longer sessions and smaller bills
Tool-mode filtering
Carries 70+ tool JSON objects with each request, keeping relevant tools and replacing the rest with stubs
4B model file compression
Compresses file reads and outputs while preserving identifiers, paths, and error messages
Original content recovery
Retrieves original bytes locally when the agent needs them, without consuming extra conversation turns
Pros
- Non-destructive compression, with original bytes retrievable locally on demand
- Simple integration: set only the ANTHROPIC_BASE_URL environment variable
- Works with Claude Code, Cursor, Codex, OpenHands, and OpenAI-compatible upstreams
- The entire stack is Apache-2.0 open source and can be self-hosted free on your own hardware
Pricing Plans
Free
- Gateway and 4B model are both open source
- Q4 quantization is about 2.5GB and runs with 8GB VRAM
- Tool filtering runs on CPU and does not require a GPU
- GitHub and Discord support
Managed
- Hosted, always-on endpoint
- No need to buy or rent a GPU; about 5 times the speed of RTX 4060
- Usage dashboard
- No credit card required
- GitHub and Discord support
* Pricing information is for reference only. Please refer to the latest announcement on Paritok's official website.