Snout

Context gating and token optimization for coding agents. Everything your agent reads stays in its context and is paid for again on every message: long markdown docs, logs, command output, MCP results, files it already read. Snout gates what goes in, so the agent gets what it needs, sessions last longer and every task costs less. Your own code always comes through in full.

npm install -g usesnout
snout scan              # how much of this repo would crowd your agent's context

Free and local: no account, and your code never leaves your machine. Works with Claude Code, Codex, Cursor, Gemini CLI and any MCP agent.

Install

# Claude Code
claude plugin marketplace add biffbuster/snout-context
claude plugin install snout@snout-context

# Codex, Cursor, Gemini CLI
npm install -g usesnout
snout init codex        # or cursor, gemini

Snout starts in observe mode and only records. After your first session it tells you what it would have saved; snout mode enforce turns it on.

How it works

Before your agent reads a file or sees a tool's output, Snout checks what it is, locally, in under a millisecond:

| What enters context | What the agent gets | | --- | --- | | Your code, config and instruction files (CLAUDE.md, AGENTS.md, skills) | Everything, in full | | Long markdown docs, specs and changelogs | The opening and a map of its sections with line numbers; any section is one read away | | Long logs | The last lines and where earlier errors are | | Something it already read, unchanged (a file, a cat, an identical MCP or command result) | A pointer to the copy already in its context | | Results from any MCP server (issue lists, wiki pages, browser snapshots, query rows) | The structure and first items, per server; the full result is saved | | Output of a passing test, install, build or search | The problems and the summary, grouped by file; the full log is saved | | Generated, vendored and built files | The first lines, an outline, and a search hint | | .env and key files | Snout asks you first |

When no rule is sure, the file is read in full, and every error lets the read through. Any ranged read (a specific set of lines) comes back exactly as asked.

Results

Real agent sessions, Snout off vs. on, each task graded by its own test. Method and caveats.

| | Result | | --- | --- | | Long multi-step sessions with docs, logs and MCP connectors (Haiku, 11 tasks, 1 run each) | −41% input tokens, −28% spend, 11/11 passing | | Spend on tasks that meet bulk files, everything on (Haiku, 48 runs) | −33%, 24/24 passing | | Spend on bulk-file tasks, read gate alone | −23% Sonnet · −24% Haiku · −23% Codex | | A large generated file · a passing test run · a large MCP result | −90% · −87% · −74% | | Tasks with nothing to trim | no change |

CLI

snout                   # status: what's been kept out, and what it saved
snout scan              # what in this repo would crowd your agent's context
snout audit             # MCP servers: which are used, what each one's results cost and saved
snout report            # what was read, kept out and saved, with spend
snout dashboard         # live savings in your browser
snout mode enforce      # start trimming (observe = record only)
snout explain <file>    # why a file was trimmed; `snout allow <file>` keeps it whole
snout login             # connect to Snout Cloud

In Claude Code: /snout:report, /snout:mode, /snout:explain. More commands: docs/configuration.md.

Snout Cloud

One dashboard for your team's agents: savings and real spend by person, project, coding agent and model. Cloud agents (CI, Claude Code on the web, Codex cloud) connect with a workspace key set as SNOUT_TOKEN. Only per-day totals are sent, never code; snout sync --dry-run shows exactly what.

| Plan | Price | Includes | | --- | --- | --- | | Free | $0 | Everything local, unlimited. Cloud dashboard for you or a team of up to 10: all projects, 90 days of history, 3 cloud-agent keys, 500 agent runs a day | | Team | $8 per developer / month, billed yearly ($10 monthly) | For larger teams: unlimited people, a year of history, unlimited keys, 5,000 agent runs a month included, then $5 per 10,000 | | Enterprise | Contact us | SSO, audit log, self-hosting |

Creating a team is invite-only while in beta. Sign in at app.usesnout.xyz.

Privacy

Snout reads a file's path, size and first 2 KB to classify it, in memory, and records its decisions to .snout/ in your project as plain files you can read or delete. It makes no network calls unless you run snout login. Releases are built and published from GitHub Actions with npm provenance. Details: SECURITY.md · docs/privacy.md.

Docs

| | | | --- | --- | | Quickstart | First session, step by step | | Integrations | Setup for each agent and MCP | | Configuration | Modes, allow lists, advanced commands | | Teams | Snout Cloud, organizations, cloud agents | | Benchmark | How the results were measured | | Troubleshooting | When something doesn't look right | | Development | Repo layout, evals, limits, contributing |

Uninstall

snout mode observe && snout logout
claude plugin uninstall snout@snout-context
npm uninstall -g usesnout
rm -rf .snout ~/.config/snout

MIT licensed.