TokenEyes
GitHub

Offline-first AI request analysis

Know what your prompt really costs.

TokenEyes inspects text, code, images, audio, and documents—then compares token usage, context fit, and API cost across OpenAI, Claude, and Gemini. Nothing leaves your machine.

Get started
01 Local by default02 Three providers03 CI-ready
~/projectLOCAL

$ tokeneyes compare --preset changed \\

--models gpt-5.5,claude-sonnet-4-6,gemini-3.5-flash

Scanning 18 sources84.2 KB
gpt-5.512,481 tok
1.19% context$0.151 expected
claude-sonnet-4-614,105 tok
1.41% context$0.102 expected
gemini-3.5-flash13,872 tok
1.32% context$0.024 expected

✓ Estimated locally · source content not saved

The full picture

One request.
Every dimension.

Model dashboards show the bill after the call. TokenEyes shows the shape of your request before it happens.

01 / COMPARE

Models speak different token.

Compare bounded estimates with provider-specific formulas, capability status, and the exact catalog snapshot behind every result.

02 / PRIVACY

Local means local.

Content stays on your machine unless you explicitly add --verify. Saved history stores hashes and counts—never source contents.

03 / PLAN

See the cliff before you hit it.

Model context fit, fixed-point cost scenarios, cached tokens, reasoning, and output budgets together.

04 / AUTOMATE

Guardrails your CI can read.

Stable JSON, deterministic scans, and distinct exit codes for context overflow, budget limits, incomplete scans, and verification failures.

--max-cost-usd 0.50exit 5

From repo to answer

Three moves.
Zero guesswork.

  1. 01

    Point

    Pass files, directories, globs, stdin, or a Git-aware preset like changed.

    tokeneyes compare .
  2. 02

    Inspect

    TokenEyes detects formats by content, applies ignore rules, and assembles the request locally.

    .go.png.pdf.wav
  3. 03

    Decide

    Compare context, cost, and support before choosing a model—or enforce a limit in CI.

    Read compare docs →
“
Every model has a context window. TokenEyes gives you a window into the context.
3provider families
15+recognized formats
0uploads by default

Open source · built in Go

Stop estimating
in the dark.

Install TokenEyes and see the request before the model does.

Read the quickstart →View on GitHub ↗