HalluciGuard reads every reply from Claude Code, Gemini CLI and GitHub Copilot. It flags doubt, backtracking and admitted mistakes in under a millisecond, before you run a single line.
The demo runs a small subset of the pattern library locally in your browser. Nothing you type leaves this page.
Dedicated pattern sets for
Code can compile perfectly and still come from an assistant that was guessing. HalluciGuard reads the prose around the code and runs the same five-step pass on every reply.
Capture the reply at the session file, API, network or library call.
Separate the assistant's prose from code blocks and user prompts.
Run compiled patterns in linear time on the Rust regex engine.
Group matches by type and rank severity from Low to Critical.
Notify by desktop, terminal, log or CLI exit code, with the exact phrase.
Pick the interception point that suits your security policy. Every mode shares the same pattern library and severity rules, so results stay consistent.
A native Node.js module built with NAPI-RS. Your agent checks its own reply and receives a halt or continue signal before it writes files or retries, which stops token-wasting loops at the source.
An OpenAI-compatible endpoint that forwards each request, checks the reply and prepends a warning before it reaches the client. Streaming supported.
Inspects TLS traffic and reassembles streamed SSE chunks. No code changes.
Reads JSONL logs, rebuilds the message tree and splits chats at each /clear. Air-gap friendly.
Scans months of history into CSV reports per file and chat for audits and trend tracking.
An OTLP gRPC receiver that plugs into your existing observability stack, ready wherever an assistant exports conversation content through telemetry.
Every pattern comes from manual review of 306+ production Claude Code sessions, Gemini CLI transcripts and Copilot community reports. Teams can extend the YAML library with their own phrases.
The assistant admits an earlier answer was wrong.
Sorry for the confusionMy mistakeI was wrong aboutI apologize forTests are edited or skipped instead of fixing code.
skip this testcomment out the assertionupdate the expected valueAn earlier recommendation is reversed mid-conversation.
Actually, let me tryWait, I need to reconsiderLet me correct thatThe same fix keeps failing in a costly loop.
still doesn't workerror persistsattempt failedThe conversation is long and details are slipping.
I need to be brieflet me summarizetoken usage warningThe assistant is unsure of its own answer.
I thinkmight workI'm not entirely surehopefullyBorderline match. Logged quietly for review.
One uncertainty or hedging phrase in a reply.
Two or more doubts, or any backtracking, failure loop or context warning.
Any error acknowledgement or test manipulation. Review before running.
| Aspect | HalluciGuard | Linters and compilers | ML-based classifiers |
|---|---|---|---|
| Looks at | The assistant's language | Code syntax and types | Generated text, scored by a model |
| Method | Deterministic patterns | Static and dynamic analysis | Neural classification |
| Latency | Under 1 ms per message | Varies by tool | Typically tens to hundreds of ms |
| Explanation | The exact phrase matched | Rule ID and line number | A probability score |
| Needs | A curated pattern library | Language config | Model weights, often a GPU |
# Developer: How do I parse JSON in Python? Actually, let me reconsider. I think you should use json.loads(), but I'm not entirely sure it handles all edge cases. This might work: import json data = json.loads(json_string) Sorry for any confusion - JSON parsing can be tricky.
The code here happens to be correct. The alert asks you to verify it before running, it does not claim the code is wrong.
Still curious? Ask us anything in the pilot form below.
No. An alert means the assistant showed signs of doubt or confusion in that reply. It is a prompt to review before you run. Confident replies can still be wrong, and hesitant ones can be right.
A single mild phrase scores Medium at most and is logged without interrupting you. Alerts escalate only when phrases stack up or the assistant admits an error. On production sessions the measured false positive rate was below 5%.
No. Detection runs locally or inside your own infrastructure. The file, batch and library modes need no network access at all.
Claude Code, Google Gemini CLI and GitHub Copilot have dedicated pattern sets. The API and network proxies can monitor any OpenAI-compatible service using the shared patterns.
Patterns are fast, repeatable and transparent. You see the exact phrase behind every alert, results are identical on every run, and there is no API key, GPU or model download.
Yes. Patterns live in YAML files, so your team can add phrases, tune severities and version the library alongside your code.
We run a batch report on your existing session history so you know exactly what HalluciGuard would flag before you switch on live monitoring.