Knapsack helps AI coders waste less context by skipping repeated files, logs, and test output.
Show changes. Skip repeats. Save tokens. Stop paying your AI coder to read the same context twice.
Install · What it does · Check it's working · Troubleshooting · Uninstall
Two paths, both measured on this repo:
| Surface | Raw | Shown to the model | Reduction |
|---|---|---|---|
| Input — reading a 17 KB source file | 6,043 tok | 3,936 tok | −35 % |
| Output — wrapping a 200-line test log | 674 tok | 80 tok | −88 % |
The compact view is real text — what the model sees is something like:
line 1
line 2
…
line 5
[Knapsack: 184 lines elided · ~628 tok · recall ks2_1b30…]
line 196
…
line 200
[knapsack 674->80 tok (-88%) · 34 blocks · 0 unchanged · 0 re-sent]
The original is one knapsack expand ks2_1b30… away — byte-for-byte. Your numbers depend on your workload; repeated reads and repeated commands benefit most.
AI coding agents read the same files and command output many times during edit-test loops. Every time, the full content gets dumped back into the context window — burning tokens on text the agent has already seen.
Knapsack notices repeated context, shows what changed, and keeps the original available for recall.
Without Knapsack:
- run tests
- AI reads a long test log
- make a small edit
- run tests again
- AI reads almost the same log again
With Knapsack:
- the first run shows the useful output
- later runs show mostly what changed
- repeated parts become short recall notes
If the agent needs the original back, it can fetch it any time — byte for byte, the same characters you'd see without Knapsack.
- repeated test output
- build logs
- search results
- large file reads
- repeated file reads
- long AI coding sessions
- keeping context cleaner
- seeing token savings
Knapsack reduces repeated context — it does not make every interaction cheaper. Savings depend on your workflow. Repeated test runs, logs, search output, and large or repeated file reads benefit most. When a shorter view wouldn't actually help, Knapsack passes the content through unchanged.
One command. Restart Claude Code. Done.
Windows (PowerShell)
irm https://raw.githubusercontent.com/MerlijnW70/knapsack/main/install.ps1 | iexmacOS / Linux
curl -fsSL https://raw.githubusercontent.com/MerlijnW70/knapsack/main/install.sh | shThe installer downloads a small binary, verifies its checksum, backs up your Claude Code config, wires Knapsack in, and runs a self-check.
After installing, restart Claude Code. That's it.
Prefer to build it yourself?
git clone https://github.com/MerlijnW70/knapsack
cd knapsack
cargo build --release
./target/release/knapsack installknapsack doctorChecks installation health: the binary, the Claude Code hook, recall storage, and a quick round-trip test.
knapsack statusShows whether Knapsack is active and how much it has saved so far.
You can also type /knapsack inside Claude Code for the same summary.
After install, you don't need to change your workflow. Run Claude Code normally — Knapsack works in the background.
Install once. Use Claude Code normally. See savings.
Right after install, knapsack status shows zero — nothing has happened yet. As soon as Claude reads a qualifying file or runs a wrapped command, savings start to accumulate.
Fastest way to see a real number — after Claude reads one biggish source file, run:
knapsack why-last 5You'll see lines like redirect-emitted 17038B H:\...\regex.rs 6043->3936 tok — the exact before/after for each Read decision, so you can see the input-reduction gain on your own files immediately.
For the cumulative picture, run knapsack status (or type /knapsack in Claude Code) for a one-screen summary, or knapsack metrics for the full scoreboard.
If Knapsack shows a shorter view, the original is still available. Claude can ask for it back any time. You usually don't need to think about this.
Power users can recall content directly:
knapsack expand <handle> # full original bytes
knapsack expand <handle> --lines 10-40 # a slice
knapsack inspect <handle> # metadata + previewknapsack doctorChecks that the hook, recall tools, and storage are healthy.
knapsack why-lastShows the last few decisions Knapsack made on file reads — useful if you expected a specific file to be shortened but it wasn't.
knapsack install --repairRe-points Claude Code at the installed Knapsack binary. Use this if you reinstalled, moved the binary, or doctor reports drift.
knapsack uninstallRemoves Knapsack from Claude Code's config. Your local Knapsack data is kept in case you reinstall.
knapsack uninstall --purgeAlso removes the local Knapsack data and cache.
Knapsack is for developers using AI coding agents — especially when those agents run tests, searches, and file reads repeatedly during edit-test loops.
Knapsack is less useful if:
- you rarely repeat commands or file reads
- your AI sessions are very short
- you don't use Claude Code-compatible hooks/MCP
Knapsack is a local developer tool. It runs on your machine, alongside Claude Code. It is not a shared or team server.
Will it lose any of my output? No. When Knapsack shows a shorter view, the original is still recoverable — same bytes, same characters.
Will it mess up my Claude Code config?
The installer backs up settings.json and ~/.claude.json before changing anything. knapsack uninstall reverses the changes cleanly.
Do I need to do anything after installing? Restart Claude Code so it picks up the new hook.
What does it need to run? Nothing extra — Knapsack is a single small binary with no runtime dependencies.
Knapsack is a small Rust binary with zero runtime dependencies. It plugs into Claude Code through its hook and recall-tool surfaces. Local data lives under ~/.knapsack/.
A deeper architecture write-up belongs in ARCHITECTURE.md — TODO.
MIT — free to use, modify, and share.