Paritok is a compression gateway that reduces token usage for AI coding agents by filtering tool schemas and summarizing conversation history.
What the product does and how it is positioned
Paritok is a compression gateway designed for AI coding agents to manage token consumption. It functions as an intermediary that filters tool schemas, compresses file reads, and summarizes conversation history to reduce the size of requests sent to upstream models.
The tool utilizes a 4B model trained on agent trajectories to identify critical code elements like identifiers and error messages. It maintains a non-destructive workflow by tagging compressed data, which enables agents to recover original content without initiating new model turns.
Source-supported ways to use the product
The company presents a customer use case where a coding agent is held to a 32,768-token ceiling, using Paritok to manage truncation and compression.
The official documentation describes using the gateway to reduce token overhead in long-running sessions with tools like Claude Code or Cursor.
The documented workflow, where available
Configure the environment variable to route agent traffic through the Paritok gateway.
Send tool schemas, conversation history, and file reads from the agent to the gateway.
Allow the gateway to filter schemas and compress file outputs using the 4B model.
Receive the processed response from the upstream model through the gateway.
Use reference tags to retrieve original, uncompressed bytes for specific segments when required.
Paritok employs three primary mechanisms to reduce the token footprint of AI agent sessions. It filters tool schemas by retaining relevant tools and stubbing others, which helps maintain cache stability across turns. The system also uses a specialized 4B model to compress file reads and tool outputs, preserving essential identifiers while reducing the overall volume of data.
Additionally, Paritok manages conversation history by summarizing turns that exceed a user-defined context budget. This approach ensures that recent turns remain untouched while older information is condensed to prevent context window overflows. The gateway architecture allows these processes to occur non-destructively, as original data remains accessible via reference tags.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
Integration is achieved by setting an environment variable, such as ANTHROPIC_BASE_URL, to point to the Paritok gateway address.
The process is non-destructive because Paritok tags compressed segments, allowing the agent to retrieve the original bytes when necessary.
Paritok works with Claude Code, Cursor, Codex, OpenHands, and any upstream model that is OpenAI-compatible.
The tool is available for self-hosting under the Apache-2.0 license and can run on hardware with at least 8GB of memory.
Once a set context budget is reached, Paritok summarizes turns that are outside the most recent window to prevent context overflow.
