Open-source · MIT · 47 probes

Is your "Claude" endpoint
actually Claude?

Cross-platform desktop app that probes any Anthropic-compatible /v1/messages endpoint to verify whether it's genuinely Claude — or a silently-swapped GPT, Qwen, DeepSeek, Llama, Gemini, or other model behind a proxy.

Download latest Try the probe editor Read the docs
47 probes · 5 categories · 3 platforms · MIT licensed

What it catches

Five categories, 47 carefully-tuned probes. Each one targets a specific behavior that's hard to fake.

Identity

Who the endpoint claims to be, knowledge cutoff, creator, training methodology, and whether it corrects false claims like "I am ChatGPT…".

Jailbreak resistance

DAN, "ignore previous instructions", grandma exploit, base64-encoded request, evil-twin roleplay.

Politically sensitive (CN)

Tiananmen, Taiwan, Xinjiang, Hong Kong, Xi, Falun Gong, Pooh — TH/ZH/EN. Catches Chinese-trained models behind a proxy.

Social-engineering / leak

Flattery + framework probe, system-prompt leak, infra probe, implicit-owner traps, forced multiple-choice, sentence-completion, wrapper trap.

Capability sanity

Math chain-of-thought, strict JSON, Thai language, Unicode round-trip, long-context needle-in-haystack.

Golden similarity

Each response is compared against 5 reference samples per probe via character-bigram cosine similarity. Language-agnostic, no tokenizer needed.

How it works

1:1 traffic emulation

Every request matches real claude-cli traffic: User-Agent, X-Stainless-*, anthropic-beta, system block, full tools array, and sha256(api_key) in metadata.

Sandbox auto-fallback

If the model replies with only a tool_use turn, the chat IPC silently retries without tools so the verdict has usable text to match.

Parallel + resumable

Runs 4 probes in flight. Stop mid-run → press ↻ Resume to continue only the remaining. Final report combines everything.

Auto-finish

When a run finishes, the Markdown report is silently saved to reports/ and published to a paste host. URL on clipboard.

Probe pack import/export

Share probe sets as portable JSON. Anyone can drop a pack into the app — no fork, no PR, no build step required.

Latency + shape fingerprint

Coefficient-of-variation flags uniform proxies. Avg chars-per-output-token and stop_reason mix add a signal beyond keywords.

Download

Pre-built binaries on every release. Source build is also one command (./run.sh or run.bat).

macOS

arm64.dmg · x64.dmg Download .dmg

Windows

win-x64.exe (NSIS) Download .exe

Linux

AppImage · deb · tar.gz Download

First-run setup

macOS — after dragging the app into /Applications/, open Terminal and run once:
xattr -cr "/Applications/Claude Verifier.app"
Windows — SmartScreen will warn "Windows protected your PC". Click More info → Run anyway.
Linux — make the AppImage executable: chmod +x Claude*.AppImage. The .deb installs with sudo dpkg -i Claude*.deb.

Both macOS and Windows show these warnings for any open-source desktop app that isn't signed with a paid Apple / Microsoft developer certificate. The workarounds are one-time per install.

Share probes with the community

The project gets stronger every time someone adds a high-signal probe. Anyone who's caught a sneaky proxy in the wild can package the prompt that caught it and drop it back into the app — no fork, no PR, no build.

Export from the app

Tick the probes you want to ship → catalog header → ⬇ Export. Save the JSON. Share anywhere — GitHub Gist, Discussions, your blog, Discord.

Import from anyone

Click ⬆ Import, pick someone else's pack. New probes appear in your catalog with a custom badge and persist across restarts.

🧪  Open the probe editor