Make coding agents resend less

Install · How it works · What you get · Proof · Supported tools · Plans · Privacy

---

Compaction runs underneath Claude Code, Codex, and Cursor. No new editor. No new agent.

It reduces eligible model visible input before it reaches the provider, shape unnecessary output before generation, and show the result inside the tools you already use.

This is a real acceptance run through a normal Codex subscription session:

$ codex

↳ compaction · input 8,388,356→7,212,095 (−14%) · output 14,393→10,795 (−25.0%, est.) · +~3.08m

The input arrow is measured before→after evidence. The output before value is a calibrated counterfactual, so it stays marked est.. +~3.08m is the estimated equivalent active agent time preserved at the observed workload consumption rate. This is not a claim about a provider's hidden quota or rate limit formula.

• Output reduction is applied from the first eligible run. No account required
• Input reduction is separately supported with a free account. We have validated ~48–50% less billed input on uncached API sessions and ~5–10% on cached sessions (provider reported)
• Local by default. Compaction processes request content on your machine and sends it only to the provider you already use. It does not upload prompts, code, or responses to any outside service, and never reads, stores, or logs your provider key

⭐ If Compaction helps, give the repo a star.

Install

curl -fsSL https://cli.compaction.dev/install | sh

Then run:

compaction

Guided onboarding detects Claude Code, Codex, and Cursor, shows what will change, and writes nothing until you confirm.

Or install from npm:

npm install -g @compaction/cli
compaction

Supported persistent installs use managed updates to keep your version up to date.

curl -fsSL https://cli.compaction.dev/install | less
curl -fsSL https://cli.compaction.dev/install | sh -s -- --dry-run

npx @compaction/cli init
npx @compaction/cli --help

compaction update --check
compaction update
compaction update --rollback

--channel preview selects npm's next tag; stable uses latest. Opt out with compaction update --auto off or COMPACTION_AUTO_UPDATE=0.

How it works

1. Install once. Run compaction
2. Connect your tools. Detection is read-only; confirmation is the first write
3. Keep working exactly as before. claude, codex, and Cursor continue to be the tools you use
4. Compaction optimizes locally. Output reduction starts on eligible runs. Live history input compaction is supported with a free account
5. Compaction writes content free receipts. Token/cache counts and structural labels only, never your prompt, code, or response
6. You can inspect the result live or later.
   • compaction watch follows new measurable turns live; --once shows the latest few and exits.
   • compaction activity shows recent run history from the local metrics-only activity store, with filters and JSON output.
   • compaction status shows setup/readiness state and includes a small recent-turn summary.

Two ways to connect

Subscription. No API key needed. Claude Code and Codex keep using the subscription you already pay for. On supported runs, subscription sessions can use both input reduction and output reduction.

API-key. Traffic goes through the local Gateway. Your key rides straight through to your provider and is never read, stored, or logged by Compaction.

Output shaping is on by default once a tool is connected. Input compaction is explicit and gated. Unsupported request shapes pass through unchanged, and the original request is retained locally for byte exact recovery.

What you get

• Output token reduction. Compaction attaches the shaping policy before eligible generation. If applicable calibration exists, it shows output ESTIMATED_BEFORE→ACTUAL (−N%, est.); otherwise we show output N/A→ACTUAL (N/A%, est.)
• Input token reduction with a free account. Compaction compresses eligible older history before it is re-sent on supported runs. Output reduction is never metered
• Content free receipts. Counts, structural labels, sources, and a short receipt id; never your prompt, code, or response
• Byte exact recovery. Every mutated request retains the original locally
• One product surface. We keep the user facing model simple: input reduction, output reduction, and evidence

Proof

Compaction only shows a number when it has evidence for it.

How Compaction reports savings

What appears depends on what Compaction could actually measure on that run.

Compaction can reduce output tokens but does not change the input:

compaction · observed input 91,472 · output 857→463 (−46.0%, est.) · id 5f539978

Compaction can also reduce model-visible input:

compaction · input 91,472→74,769 (−18%) · output 857→463 (−46.0%, est.) · 1.92M/2M left · id 5f539978

On an API-key, where the provider bills at a published per token rate, Compaction can also show the estimated price impact:

compaction · input 91,472→74,769 (−18%) · output 857→463 (−46.0%, est.) · −$0.05 (list price) · 1.92M/2M left · id 5f539978

We use a simple evidence rule:

• Input before→after is measured. Both values exist for the same request
• Actual output is measured.
• Output before is estimated. The same generation cannot simultaneously happen with and without Compaction, so the counterfactual remains marked est.
• If Compaction cannot support a number with evidence, it is shown N/A.

For example:

compaction · observed input 91,472 · output N/A→463 (N/A%, est.) · id 5f539978

This is the beginning of the README. Read the rest at the original. Read the full README ↗