Context compression for coding agents: approaches compared

There are now many tools that promise to shrink what a coding agent reads. They work at very different points: before a command runs, between the agent and an MCP server, on the whole request on the wire, or on one result against the current question. Here is how the approaches compare and how to test any of them on your own work.

October 9, 2026 · Comparison

In short: pick by where your tokens go. Noisy shell commands: an output filter. Huge MCP results: a proxy or result compression. Long sessions: built-in compaction. Bulky output where only a few lines answer the current question: query-aware extraction such as SuperCompress. Measure on your own sessions, not on a vendor headline.

The approaches

Approach	Examples	Where it acts	Good fit	Watch out for
Command output filters	RTK, chop	Rewrites or filters known shell commands before the agent reads them.	Noisy `git`, test, build and install output.	Only known command formats; file reads and MCP results pass through.
MCP proxies with retrieval	sqz, mcp_trunc_proxy, context-compress	Between the agent and MCP servers or the shell; keep the full result aside and return a preview.	Very large MCP results.	The agent has to know to fetch more.
Wire-level proxies	Headroom	Rewrites the whole request between the agent and the model provider.	Your own apps and pipelines.	Changes model routing; rewriting history can affect prompt caching.
Built-in compaction	`/compact` in Claude Code, Codex and OpenCode; `/compress` in Gemini CLI	Summarizes the whole session once it is large.	Long sessions.	Summaries can lose exact details.
Query-aware extraction	SuperCompress	One bulky result at a time, scored against the current question; keeps original lines.	Logs, diffs, file reads and MCP results where a few lines matter.	Files the agent is about to edit should stay whole.

How to measure any of them

Measure what you are billed, per finished task. Fewer tokens in one request does not always mean a cheaper session. Independent work on coding-agent compression, such as the arXiv paper An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents and the open tokbench harness, makes this point.
Watch prompt caching. Rewriting earlier turns can break cached prefixes. SuperCompress compresses new bulky results and leaves earlier turns alone, which is the cache-friendly pattern.
Gate on correctness. Count a session only if the task was done right.
Run your own repo. Every SuperCompress call reports original_tokens and kept_tokens, so you can see each call on your own work.

How SuperCompress works

SuperCompress is hosted, extractive and query-aware. Your agent sends a large block of context (a log, a diff, a file read, an MCP result) together with the question it is working on. Our Neural Keep model scores the lines against that question, keeps the original lines that answer it, and drops the rest. It never rewrites or summarizes, so file paths, error messages, IDs and code come back exactly as they were. The question itself is never compressed.

Every compress call reports original_tokens and kept_tokens, so you can see what was dropped on your own work instead of trusting a headline number. The usage_summary MCP tool shows totals per agent for your account.

Install

npm install -g supercompress-proxy && supercompress setup

Setup links your SuperCompress account in the browser, detects the coding agents on the machine and wires each one. Restart the agent so it loads the new MCP server. Then check what is wired up:

supercompress doctor

doctor prints a per-agent health check. supercompress uninstall removes only the entries SuperCompress added. Setup backs up each config file before it edits it.

Head-to-head pages: SuperCompress vs RTK, SuperCompress vs Headroom, vs summarization, vs truncation.

FAQ

What is context compression for coding agents?

Reducing what a coding agent sends to the model, usually by shrinking tool output, file reads and MCP results, or by compacting the session.

What is the best context compression tool for Claude Code?

It depends on where your tokens go. Command filters help with noisy shell output, MCP proxies with huge MCP results, built-in /compact with long sessions, and SuperCompress with bulky output where only a few lines answer the current question. Measure on your own sessions.

Is SuperCompress a summarizer?

No. It is extractive: it keeps original lines and drops the rest, so error messages, paths and code are unchanged.

Can I use SuperCompress with RTK or built-in compaction?

Yes. They act at different points and can be combined.

Get API key Coding agents guide All coding agents