A summarizer produces new text you then have to trust; this removes tokens and leaves the rest in place, which is why it can be pointed at logs, issue threads and research dumps without changing what they say. The refusal list is the good part of the design — diffs, lockfiles, SQL and `.env` files are declined rather than mangled, and `auto` mode leaves code intact while squeezing the comments around it.
An MCP adapter over ContextCrumb, a compressor that works word by word rather than by rewriting. It is not a summarizer — surviving text stays in source order with expendable tokens removed, so names, actions, constraints and sequence stay put while filler goes.
- Compress inline text with a keep-probability `threshold`, or a fixed `target_keep_ratio` when the budget rather than the cutoff is the constraint — `compress_text`
- Compress a file on the machine running the server, with `encoding` and the same threshold controls — `compress_file`
- Code-aware handling in `auto` mode: executable code is preserved exactly and only comments and docstrings are compressed, for Python, JavaScript, TypeScript, JSX, TSX, Go and Rust
- `content_mode` to choose the behaviour yourself — `auto`, `prose`, `code-comments`, `raw` or `refuse`
- A refusal path for syntax-sensitive files — diffs, structured configs, lockfiles, SQL and `.env` — unless you pass `force`
- `return_tokens` to see which tokens were kept and which were dropped
- A separate proxy, `contextcrumb-shrink`, that compresses verbose catalog descriptions before an agent reads them while passing tool names, schemas, calls and results through unchanged
Python >=3.10 and the MCP extra: `pip install "contextcrumb[mcp]"`, or run it with no install at all via `uvx --from "contextcrumb[mcp]" contextcrumb-mcp` — which is exactly what the documented client entry does. `compress_file` reads paths on the server's machine, not the client's. For repeated calls, start `contextcrumb service start --allow-root .` and add `--use-service` so each server process is not loading its own model.
One command — pip install "contextcrumb[mcp]"