Today we're launching Lineman, an AI sidekick for Claude Code that takes on the heavy lifting your main model shouldn't have to. It's available now as a plugin. Setup takes a couple of minutes, and it starts paying for itself on the first big file you read.
Coding agents burn most of their context window on data, not reasoning. Picture a 1,200-line file you needed three functions from, a 400-line test log where one assertion failed, or a web page wrapped in 8KB of navigation chrome. All of those tokens crowd out the thinking that actually solves your problem, and you pay for every one. Lineman fixes that.
What Lineman does
Lineman is an MCP server that sits between Claude Code and your data-heavy tools. When your agent reads a big file, triages a build, greps a codebase, or fetches a web page, Lineman catches the result and hands it to a fast secondary model. That model compresses it down to what matters before it reaches your primary model.
Your main context stays focused on the reasoning. The token-heavy grunt work happens on cheaper infrastructure. You get the same answer for a fraction of the spend.
Keep the expensive model thinking. Send everything else to the sidekick.
The savings
We measured Lineman against 180 tasks across six benchmark suites, including real-world GitHub issues from SWE-bench. The headline number is a 53% average reduction in tokens, with output quality essentially flat.
| Workload | Token reduction |
|---|---|
| SWE-bench Lite (real GitHub issues) | 48% |
| Average across 6 benchmark suites (180 tasks) | 53% |
| Internal data-heavy tasks | up to 75% |
That saving doesn't come out of your accuracy. Across external benchmarks, Lineman held 98.3% of baseline output quality. You're cutting the data tax, not the intelligence.
How it works
Lineman uses summary-first interception. Hooks in Claude Code catch native tool calls (Read, Bash, Grep, Glob, WebFetch) and route their output through compression:
- Large files come back as a structured summary with the symbols, signatures, and sections your agent asked about, not 1,200 verbatim lines.
- Build and test output is triaged down to the failures and what caused them.
- Search results are re-ranked and condensed instead of dumped raw.
- Web pages are stripped to their substantive content.
Anything not worth compressing passes straight through. The model-facing surface stays deliberately small: two tools, assist and edit_file. There's nothing new to learn. You keep using Claude Code the way you already do, and Lineman works in the background.
Pricing
Lineman's pricing is anchored to the tokens you save, not the tokens you spend. Every plan starts with a 14-day free trial, no credit card required.
| Plan | Price | Monthly tokens |
|---|---|---|
| Basic | $49 / month | 100M |
| Pro | $149 / month | 440M |
| Enterprise | Custom | Unlimited |
See the pricing page for the full breakdown and annual options.
Getting started
Lineman installs as a Claude Code plugin. Inside a Claude Code session:
/plugin marketplace add lineman-io/lineman-mono
/plugin install lineman@lineman
/reload-plugins
/lineman:auth
That's the whole setup. /lineman:auth opens your browser once to sign in, and your next large file read goes through Lineman automatically. Run /lineman:stats whenever you want to see how many tokens you've saved.
What's next
This is just the start. We're expanding the set of task types Lineman can compress, bringing latency on the secondary model down further, and building richer session tooling. Follow along here in News for product updates and engineering notes as they ship.