New

Snout 0.3.3 is out: your agent can search this project's history, on your machine.

New

Snout 0.3.3 is out: your agent can search this project's history, on your machine.

Better context for coding agents, on your machine.

Better context for coding agents, on your machine.

Snout keeps junk out of your coding agent’s context. In a 140-session Claude Code benchmark, cost per solved task fell 25.0% with all attempts passing.

Free and local by default. Optional network features require opt-in.

Works with

Claude Code

Codex

Tested on

TypeSafe Jev

140 benchmark sessions

Bulk-file tasks: 29.3% lower cost per solved task.

Across 100 bulk-path sessions, all attempts passed and cost per solved task fell 29.3% (95% CI: −44.3% to −11.2%).

shopfront
On

Why did the checkout test start failing?

  1. Checking each file
  2. Trimming bulky files
  3. Reading your code in full
  4. Adding up 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%
Kept outlockfilegenerated codelibrarybundleExplainView report
Ask the agent…Allow fileOpen report
shopfront
On

Why did the checkout test start failing?

  1. Checking each file
  2. Trimming bulky files
  3. Reading your code in full
  4. Adding up 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%
Kept outlockfilegenerated codelibrarybundleExplainView report
Ask the agent…Allow fileOpen report

Coding agents read everything they open, then re-send all of it with every new message. So a junk file read in minute one gets paid for again on every turn after. Snout stops the junk at the door.

01

Audit

Find and clean agent files

See every instruction file, skill and AI-written doc your agent loads, what each one adds to every session, and which are stale or unused. Archive them in one step and bring them back any time.

02

Benchmarks

Tested on real agents and workflows

Our benchmark of the 0.3.3 release ran Claude Code Sonnet on Windows: 14 tasks, five repeats per arm. All 70 attempts passed in each arm. Codex needs a matched 0.3.3 run.

03

Tokens

Fewer tokens, lower costs

Snout filters bulk reads and trims long tool output in Claude Code. In the 0.3.3 Sonnet benchmark, cost per solved task fell 25.0%. Your results will vary by task.

Less cost in this benchmark. Quality still under study.

Snout · a context gate for AI coding agentsThe agent opens five files totalling 137 thousand tokens. With Snout, the two source files go in whole and the lockfile, generated client and bundled library arrive as short summaries: 16 thousand tokens in context, 121 thousand kept out.WHAT THE AGENT OPENSsrc/cart.tsYour code6kcart.test.tsYour code4kyarn.lockDependency list61k → 2kclient.gen.tsGenerated code38k → 3kvendor/sdk.jsBundled library28k → 1kKEPT OUT121ksnout checks each fileAGENT CONTEXTWITH SNOUTYour codeSummaryIn context16kKept out121k1 SQUARE = 1K TOKENS
01 / 05

Claude Code can return bounded excerpts of classified bulk; Codex gates supported reads. Handwritten source generally stays readable.

Snout · a context gate for AI coding agentsThe agent opens five files totalling 137 thousand tokens. With Snout, the two source files go in whole and the lockfile, generated client and bundled library arrive as short summaries: 16 thousand tokens in context, 121 thousand kept out.WHAT THE AGENT OPENSsrc/cart.tsYour code6kcart.test.tsYour code4kyarn.lockDependency list61k → 2kclient.gen.tsGenerated code38k → 3kvendor/sdk.jsBundled library28k → 1kKEPT OUT121ksnoutAGENT CONTEXT16k in · 121k out1 SQUARE = 1K TOKENS
01 / 05

Claude Code can return bounded excerpts of classified bulk; Codex gates supported reads. Handwritten source generally stays readable.

What the 0.3.3 benchmark measured.

The current matched benchmark is Claude Code Sonnet on Windows. Earlier Haiku and Codex runs used older builds; we keep them as historical results.

Paired runs · Snout off vs onWhat changed with Snout on
Run
−50%−40%−30%−20%−10%0+10%
Change
Current · Claude Sonnet v0.3.3
Historical · pre-0.3.1
← Cheaper with SnoutCostlier →
Current · Claude Sonnet v0.3.3Claude Sonnet v0.3.3, every feature on
−25.0%
cost per solved task, Snout on vs off
95% range−38.2% to −9.5%
−50%0+10%
Saving provenThe whole range is below zero.
Off70/70
On70/70
Every attempt passed its tests, with Snout off and on.

Windows, 14 authored tasks with five repeats per arm. All 70 attempts passed in each arm. This does not prove no quality loss.

Passing attempts70/70attempts passed with Snout on, as did 70/70 with it off. This sample does not establish no quality loss.
Cost per solved task−25.0%in the 0.3.3 Sonnet benchmark on 14 authored tasks. Five tasks cost more with Snout.
One junk file, re-sentExample

Say your agent reads a 60k-token lockfile once. It goes out again with every message after.

Messages after the read10
Without Snout600k
With Snout60k
Tokens re-sent. In our tests a lockfile read came back about 90% smaller.

Current v0.3.1 Sonnet results come from 140 sessions on 14 authored Windows tasks. Older build results are labeled historical. Ranges are 95% task-bootstrap intervals; quality noninferiority is not established. How we measure →

Explore a measured savings scenario

Estimate from our Claude Code Sonnet test of the 0.3.3 release. Your savings may differ; Codex has not yet had a matched cost test.

Illustrative savings scenarioESTIMATE
Who's using it
Agent
Your plan
Bulky files in your project

Illustrative rate from the whole 14-task authored Sonnet suite.

Illustrative only · v0.3.3 Sonnet benchmark · not a plan-limit prediction
SUBSCRIPTION LIMIT EFFECT
Varies
API COST RESULTS DO NOT PREDICT PLAN LIMITS
BENCHMARK COST / TASK
−25.0%
TASKS STILL PASSING
70/70
same as without Snout
SNOUT COSTS
$0
free, runs locally
CURRENT EVIDENCE
Sonnet
140 benchmark sessions
1 SQUARE = 1% OF BENCHMARK COST PER SOLVED TASK25% LOWER IN BENCHMARK

Claude Code Sonnet v0.3.3 on 14 authored Windows tasks: −25.0% cost per solved task (95% CI −38.2% to −9.5%). All 70 attempts passed in each arm; five tasks cost more with Snout. Your result may differ.

Connect your agent
Illustrative savings scenarioESTIMATE
Who's using it
Agent
Your plan
Bulky files in your project

Illustrative rate from the whole 14-task authored Sonnet suite.

Illustrative only · v0.3.3 Sonnet benchmark · not a plan-limit prediction
SUBSCRIPTION LIMIT EFFECT
Varies
API COST RESULTS DO NOT PREDICT PLAN LIMITS
BENCHMARK COST / TASK
−25.0%
TASKS STILL PASSING
70/70
same as without Snout
SNOUT COSTS
$0
free, runs locally
CURRENT EVIDENCE
Sonnet
140 benchmark sessions
1 SQUARE = 1% OF BENCHMARK COST PER SOLVED TASK25% LOWER IN BENCHMARK

Claude Code Sonnet v0.3.3 on 14 authored Windows tasks: −25.0% cost per solved task (95% CI −38.2% to −9.5%). All 70 attempts passed in each arm; five tasks cost more with Snout. Your result may differ.

Connect your agent

01

SUPPORTED AGENTS

Works with your stack

Snout connects to Claude Code and Codex in your project. Follow the setup steps for your agent.

Free local core. No account required. Optional network features are opt-in.

Set it up

Private by design, honest by default

Runs on your machine

Core Snout features run locally. Optional network features require opt-in. No account is needed for local use.

When unsure, it reads

If Snout is uncertain about a file, your agent gets the full read. You can allow specific files.

You stay in control

Change modes anytime, and allow specific files to be read in full.

Install Snout

Terminal

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

Questions

Will my agent get worse?

Will my agent get worse?

Does my code leave my machine?

Does my code leave my machine?

How much will I save?

How much will I save?

Does it work with my AI tool and plan?

Does it work with my AI tool and plan?

What counts as junk?

What counts as junk?

What does it cost?

What does it cost?

Spend tokens on your code, not clutter.

Spend tokens on your code, not clutter.

Install the current free release (v0.3.3), start in observe mode, and inspect your local report after a session. Our cost benchmark measured this release.

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