Lineman runs a cheap second model as a compressor, filter and classifier. It chews through the data-heavy bytes - big files, build logs, search dumps - and hands Claude a small, structured summary. The swap happens automatically, before the tool result ever reaches Claude. You change nothing about how you work.
Lineman sits on the wire between Claude Code and its tools. The swap is automatic and invisible - you keep working exactly as you do today.
A native Read, Bash or Grep runs as usual. A Lineman hook catches the result the instant it returns - before Claude ever sees it.
The bulky output is streamed over TLS to a cheap secondary model. None of it touches your expensive Claude context.
It reads the noise for you - classifying, filtering, summarising - and keeps only what is relevant to the task at hand.
A compact, structured result lands back in Claude Code. Your context window only ever spends tokens on the signal.
Hooks intercept Claude Code's native tools and replace the bulky result with a compressed summary - before it ever reaches the model. Everything else on this page hangs off this one move.
Large reads are swapped for a summary or a structural map - not every line dumped into context.
Build runs, installs and logs are compressed down to their key findings.
Hits are re-ranked by relevance and trimmed instead of returned raw.
Long globs are grouped and condensed into a structure you can scan.
A fetched URL is reduced to the part that's actually relevant to the task.
Big images become a compact text descriptor, with the full image one call away.
Every large result is either compressed or explicitly marked as a deliberate pass-through, with a typed reason. Silent skipping is treated as a bug - so you can always trust that nothing was quietly dropped on the way to Claude.
Re-reading files is the single biggest token sink in agentic coding. Lineman climbs a ladder, cheapest rung first, and only pays for more when it has to.
Build, test and CI output is enormous and mostly noise. Lineman turns a 500-line failure into the three lines that actually matter.
Parses compiler and test-runner output and extracts the actual errors and their locations.
Classifies a crash into a typed diagnosis instead of feeding the whole trace into context.
Scans logs for the errors and warnings relevant to the issue being chased.
Condenses test runs, large diffs and git history into something readable.
Results are ordered so the most relevant hit is first, and oversized sets are trimmed.
A handful of hits? Lineman skips the cloud step and returns them raw - no point paying to rank one screen.
The classic agent loop reads a whole file just to change three lines - doubling that file's token cost. Lineman removes the round-trip.
Lineman confirms the exact target text exists before the edit and applies the change in one step - no “read the whole file, then edit it” detour.
When the target isn't found or is ambiguous, Lineman returns the candidate locations and closest match, so Claude fixes it without re-reading or re-grepping the file.
Claude Code's native auto-compaction bluntly truncates the conversation when context runs out. Lineman intervenes first - it spots completed chunks of work, moves them into recoverable storage, and leaves a compact pointer behind. Done work then costs almost no context, but stays fetchable.
A token optimiser that quietly costs a fortune would be self-defeating. Lineman keeps its own footprint honest - and shows its working.
Lineman exposes only two model-facing tools. Every tool description is re-paid on every single turn - a lean surface keeps the standing cost low.
2 tools · re-paid every turnThe tool descriptions are deliberately minimal. Trimming them alone improved a benchmark - wordy descriptions are a tax you pay on every turn.
+23% from trimming aloneLineman never injects persistent “reminders” into context. A standing reminder is a tax paid every single turn - so there are none.
0 persistent remindersEvery mechanism on this page ties back to something you can watch happen - and verify against your real Anthropic bill.
The headline number discounts what Claude's own native compaction would have done anyway - so it matches what you can verify on your bill. We never inflate with naive totals.
A per-session breakdown of tokens saved - reads, builds, search, web - feeds straight into the Cost Explorer dashboard.
Open the Cost Explorer to see savings by task type, session over session.
Open the Cost ExplorerIf you have something we haven't covered, email us and we'll get back to you.
Yes. The Free plan needs no credit card and gives you the cost and spend reporting dashboard, tracking up to $1,500 per month of AI agent spend across your repos, tickets, and models. Paid plans add Lineman's token-saving tool (40%+ fewer tokens), team seats, and uncapped reporting. You can start a 14-day trial of a paid plan at any time.
One config line. Works with Claude Code today, more clients coming soon. Most users see their first compressed call within 30 seconds.