New

Reads now get trimmed, not blocked: −24% cost on bulk-file tasks in our benchmark.

New

Reads now get trimmed, not blocked: −24% cost on bulk-file tasks in our benchmark.

Feed your agent
only what matters.

Feed your agent
only what matters.

Snout checks every file before it enters your coding agent’s context. Lockfiles, generated code, and vendored bundles get held back, and the agent gets a map of what’s inside instead.

Runs on your machine. No network calls, no API key.

Works with

Claude Code

Codex

Accuracy checked with

TypeSafe Jev

TypeSafe Jev

One lockfile. 60k tokens your agent never had to read.

The agent asked for yarn.lock. Snout sent the first 60 lines and a search hint instead, and the fix still landed.

shopfront
Enforce

Why did the checkout test start failing?

  1. Classifying each read
  2. Trimming bulk files
  3. Reading source in full
  4. Recording the savings

~140k tokens stayed out of context this session. Every source file the fix needed was read in full.

Tokens per read, with Snout vs. withoutWith SnoutWithout
yarn.lock1,850−97.0%
client.gen.ts2,900−93.0%
vendor/sdk.js1,100−96.2%
dist/app.min.js420−97.9%
cart/checkout.ts3,200+0.0%
Classified aslockfilegeneratedvendoredminifiedExplainView report
Ask the agent…Allow fileOpen report
shopfront
Enforce

Why did the checkout test start failing?

  1. Classifying each read
  2. Trimming bulk files
  3. Reading source in full
  4. Recording the savings

~140k tokens stayed out of context this session. Every source file the fix needed was read in full.

Tokens per read, with Snout vs. withoutWith SnoutWithout
yarn.lock1,850−97.0%
client.gen.ts2,900−93.0%
vendor/sdk.js1,100−96.2%
dist/app.min.js420−97.9%
cart/checkout.ts3,200+0.0%
Classified aslockfilegeneratedvendoredminifiedExplainView report
Ask the agent…Allow fileOpen report

From every read to real savings

Claude Code

hooks

Plugin

observe

Observe

Installed. Nothing blocked yet.

1

Install in one command

Add the plugin and restart Claude Code. Snout starts in observe mode: it records every read and blocks nothing.

Fix the failing checkout test.

Your prompt

Run

2

Work the way you already do

Prompt your agent as usual. Snout classifies each file it reads by name, path, and first 2 KB, on your machine, in under a millisecond.

yarn.lock · lockfile · 1.00
trimmed to lines 1–60 of 14,812
try: grep -n 'left-pad@' yarn.lock

Trimmed, not blocked

3

Bulk files get trimmed

In enforce mode a lockfile or generated client arrives cut to its head, with an outline and a search hint. Unsure means read: source always comes through in full.

File

Read

Held

yarn.lock

2k

61k

client.gen.ts

3k

38k

sdk.js

1k

28k

This session’s savings

Report

4

See what it saved

The report shows tokens read, held back, and repeated, per agent. Explain shows exactly why any file was flagged.

Most tools that hide files ask you to trust them. Snout shows its reasoning for every file: the rule that fired, the confidence, and the threshold it crossed. If it’s wrong, one command stops it for good.

The label

Lockfile, generated, vendored, minified, build output, secret, or source.

The score

A confidence from 0 to 1 for every label, not just the winner.

The band

Act at 0.90 and up, ask at 0.50, read in full below that. Both thresholds are yours to tune.

The override

Allow a path once and Snout never flags it again.

yarn.lock is a lockfile. Written by the package manager, not by hand.

Explain

Scores

Bands

Overrides

Ran on your machine

Copy

/snout:explain yarn.lock
​
label lockfile · confidence 1.00
rule lockfile (tier 0)
band act · enforce trims it, advise asks
thresholds act ≥ 0.90 · ask ≥ 0.50
​
SCORES read = 1 − the strongest flag
lockfile 1.00 ████████████████████
oversized 0.62 ████████████▍
read 0.00

Your team’s savings, every Monday.

Snout Team rolls every developer’s sessions into one weekly report: tokens held back, files agents keep re-reading, and new bulk files entering the repo. Team is in early access.

Team savings this week

Monday, 8:00 AM

Monday, 8:00 AM

Low-value reads fell 94%. Every task still passed.

Lockfiles

412k tok

Generated

238k tok

Vendored

171k tok

Files agents keep re-reading

Monday, 8:00 AM

Monday, 8:00 AM

package-lock.json was read 9 times. One line in CLAUDE.md stops it.

api

web

worker

docs

New bulk files in the repo

Monday, 8:00 AM

Monday, 8:00 AM

Two generated clients landed this week. Snout already flags both.

What Snout holds back

Rules read the name, the path, and the first 2 KB. No model, no network. Anything they aren’t sure about is read in full.

File type

Example

What the agent gets

Lockfiles

yarn.lock

The head, plus a search hint

Generated code

client.gen.ts

The head, plus an outline

Vendored code

vendor/

The head, plus an outline

Minified files

app.min.js

A bounded search hint

Build output

dist/

Held back only if .gitignore ignores it

Secrets

.env

Snout asks you first, every time

Your source

src/

Read in full

Built for the person who reviews what runs

Local by design

Rules run on your machine. No network calls, no API key, and no file contents sent anywhere.

Unsure means read

When no rule is confident, the file is read as normal. Every error path lets the read through.

Nothing changes without --yes

Suggested changes are previewed with evidence, applied only when you confirm, and undoable.

What you run once

Terminal

# add the plugin, then restart Claude Code
claude plugin marketplace add biffbuster/snout-context
claude plugin install snout@snout
# observe first; switch on when you’re ready
/snout:mode enforce
/snout:report
# add the plugin, then restart Claude Code
claude plugin marketplace add biffbuster/snout-context
claude plugin install snout@snout
# observe first; switch on when you’re ready
/snout:mode enforce
/snout:report

Questions

Will Snout hide code my agent needs?

Will Snout hide code my agent needs?

Does my code leave my machine?

Does my code leave my machine?

How much does it actually save?

How much does it actually save?

Which agents does it work with?

Which agents does it work with?

What if it flags the wrong file?

What if it flags the wrong file?

What does it cost?

What does it cost?

Give your agent less to read.

Give your agent less to read.

Install in one command, see what it would hold back, then switch it on.

Snout, a pink anteater in a cowboy hat
Snout, a pink anteater in a cowboy hat