claw-tsaver is an MCP proxy tool designed for OpenClaw users, with the core goal of significantly reducing the token consumption generated by tool calls through a lazy-loading extension mechanism.
Purpose MCP tool calls (such as fetch and puppeteer) often return thousands of tokens of HTML or JSON in a single response, while the model typically uses only about 5% of that content; the remaining 95% silently consumes the context window and increases usage costs. As a proxy layer between OpenClaw and downstream MCP servers, claw-tsaver is specifically designed to solve this problem.
Main capabilities
- Intercept oversized responses: intercept when a response exceeds a configurable token threshold.
- Store the full content in a local SQLite database.
- Return a compact preview and an on-demand expansion handle to the model.
- When the model needs more content, it can call the expand_content tool to retrieve it on demand.
- It is officially claimed to reduce tool call payloads by more than 90%.
- Provide an optional local Web UI for viewing token savings statistics in real time.
How to use
- Install uv (one-time setup). No need to install claw-tsaver; uvx will fetch and run it on demand.
- Configure the downstream MCP server: edit ~/.claw-tsaver/config.json (running claw-tsaver-mcp for the first time will automatically create a template).
- Register with OpenClaw: add the corresponding configuration block at the top level of ~/.openclaw/openclaw.json, then restart the OpenClaw gateway: openclaw gateway restart.
- Optional: start the local Web UI to view real-time token savings statistics, and open it in the browser at http://localhost:7878。
Module composition
- Module A: Lazy-loading extension proxy (released in this version).
- Module D: Local dashboard (released in this version).
- Module B: Tool routing (automatically load only relevant MCPs each round, not yet released).
- Module C: Conversation history compression (atomic fact cards, not yet released).
Measured data It is officially claimed to have been tested on 2026-04-25 on OpenClaw + Claude Sonnet 4.6 + mcp-server-fetch; the raw data is available in benchmarks/mvp-day1-fetch.jsonl.
Applicable scenarios Suitable for users who use OpenClaw and have integrated MCP servers that return large amounts of HTML or JSON (such as fetch and puppeteer), and who want to reduce context window usage and lower token costs.
Limitations
- The specific configurable token threshold range is not yet provided.
- The features of Module B and Module C have not yet been released.
- Except for the measured environment above, results in other environments are not yet provided.
- The project is open source under the MIT license; see the LICENSE file for details.
