klovys99

Compatible with Claude Code, Codex, and other OpenAI-compatible clients.
Anonymizes sensitive values before they leave the machine, sends only tokenized payloads upstream, then restores the original values locally in responses.
Includes a local dashboard and test tool, powered by Presidio, Gitleaks, a local NER engine, and SQLite for stable anonymization and de-anonymization round-trips.

Architecture At A Glance

flowchart LR
    vscode["VS Code clients<br/>Claude extension / Codex extension"]
    proxy["klovys99 local proxy"]
    apis["Remote APIs<br/>Anthropic / OpenAI-compatible"]

    vscode -->|"1. raw prompt / request"| proxy
    proxy -->|"2. anonymized request only"| apis
    apis -->|"3. provider response"| proxy
    proxy -->|"4. restored response"| vscode

    classDef klovys fill:#1371E6,stroke:#1371E6,color:#ffffff;
    class proxy klovys;

flowchart LR
    client["Local client or web UI"]
    input["Prompt / tool output / text payload"]
    engine["klovys99 anonymization engine"]
    regex["Regex detectors"]
    presidio["Presidio rules"]
    gitleaks["Gitleaks rules"]
    ner["Open-source NER"]
    sqlite["SQLite mapping store"]
    upstream["LLM provider"]
    restore["Response de-anonymization"]

    client --> input
    input --> engine
    engine --> regex
    engine --> presidio
    engine --> gitleaks
    engine --> ner
    engine <-->|"stable token map"| sqlite
    engine -->|"anonymized text only"| upstream
    upstream -->|"tokenized response"| restore
    sqlite --> restore
    restore -->|"clear response restored locally"| client

    classDef klovys fill:#1371E6,stroke:#1371E6,color:#ffffff;
    class engine,restore klovys;

flowchart LR
    runtime["klovys99 process"]
    admin["Admin web UI<br/>dashboard, stats, controls, test tool"]
    chat["User web UI<br/>Gemini / Claude / ChatGPT / Mistral AI"]
    auth["OAuth login<br/>work in progress"]
    keys["Current auth mode<br/>user enters API keys"]

    runtime --> admin
    runtime --> chat
    chat --> keys
    chat -. planned .-> auth

    classDef klovys fill:#1371E6,stroke:#1371E6,color:#ffffff;
    class runtime klovys;

Get Started

npm install is currently broken. Until it is fixed, run klovys99 from source: start the app manually with Go and start the GLiNER sidecar with Docker, as described below.

The simplest way to start everything (GLiNER sidecar + proxy) in one command:

./scripts/dev-start.sh

This builds the GLiNER Docker image, downloads the pinned model on first run, starts the sidecar container, and then runs go run ./cmd/klovys99.

If you prefer to run the steps yourself from the repository root:

Build and start the GLiNER sidecar with Docker.

docker build -t klovys99-gliner:local sidecar/gliner
docker run --rm --user "$(id -u):$(id -g)" \
  -e GLINER_MODEL=urchade/gliner_multi_pii-v1 \
  -e GLINER_MODEL_REVISION=1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d \
  -e GLINER_MODEL_DIR=/models/model \
  -e HOME=/tmp -e XDG_CACHE_HOME=/tmp/.cache -e HF_HOME=/tmp/.cache/huggingface \
  -v "$HOME/.klovys99/gliner:/models" \
  klovys99-gliner:local python /app/install_model.py
KLOVIS_GLINER_MODEL=urchade/gliner_multi_pii-v1 \
KLOVIS_GLINER_MODEL_REVISION=1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d \
KLOVIS_GLINER_DATA_DIR="$HOME/.klovys99/gliner" \
  docker compose -f sidecar/gliner/compose.yaml up -d --no-build

Start the local proxy with Go, pointing it at the sidecar you just started.

KLOVIS_GLINER_MODE=full \
KLOVIS_GLINER_URL=http://127.0.0.1:8091 \
KLOVIS_GLINER_MODEL=urchade/gliner_multi_pii-v1 \
KLOVIS_GLINER_MODEL_REVISION=1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d \
  go run ./cmd/klovys99

Klovys99 listens on http://127.0.0.1:8080 by default and exposes:

http://127.0.0.1:8080/anthropic for Claude Code and other Anthropic clients
http://127.0.0.1:8080/openai/v1 for Codex and other OpenAI-compatible clients

The historical unprefixed route also still exists and forwards to KLOVIS_TARGET_URL, which defaults to https://api.anthropic.com.

Optional: launch the AI workspace UI, a chat interface that anonymizes your prompt before sending it to a provider you configure with your own API key. If you want the Compte IA / AI accounts panel to persist provider API keys across reloads and restarts, set KLOVYS99_AI_WORKSPACE_KEY first (you can generate one with openssl rand -hex 32).

npm run ui:ai-workspace

This starts a local Vite dev server at http://127.0.0.1:3001 that proxies /api and /dashboard requests to the running proxy on port 8080. The dashboard header and the AI workspace link both point here. See AI Workspace for details.

If you installed the published npm package instead of cloning this repository, use the same flow with npx klovys99:

Point your client at the proxy manually — see Client Configuration below, since the configure CLI helper is part of the currently broken npm flow.

Features

Local reverse proxy for Anthropic and OpenAI-compatible JSON requests.
Built-in deterministic detectors for common PII and sensitive identifiers.
Local GLiNER Docker sidecar with explicit full and off modes.
Dynamic detector loading from the official Gitleaks and Microsoft Presidio rule sources.
Stable pseudonym tokens for the lifetime of the proxy process.
Structured logs with anonymization counters instead of raw prompt values.
Disk cache for downloaded external rules to avoid repeated network fetches on every startup.
A dashboard for live anonymization stats and a manual anonymization test tool.
An AI workspace UI for chatting with Claude, Gemini, OpenAI, or Mistral using an anonymized prompt, without leaving the anonymization boundary.

Requirements

Go 1.25 or newer.
Docker Desktop or Docker Engine, to run the GLiNER sidecar.
Network access on first startup to download the default Gitleaks and Presidio rule sources, and the pinned GLiNER model.
An Anthropic API key, Claude subscription, or OpenAI API key depending on the client you route through Klovys99.

Check your local tooling:

go version
docker version

Installation

npm installation is currently broken. Build and run from source instead:

git clone https://github.com/Korbicorp/klovys99.git
cd klovys99
go build -o klovys99 ./cmd/klovys99

See Get Started for how to run it alongside the GLiNER sidecar.

Client Configuration

The configure CLI helper is part of the currently broken npm flow. Point your client at the proxy by editing its config file directly.

Codex

Edit ~/.codex/config.toml and set:

openai_base_url = "http://127.0.0.1:8080/openai/v1"

Claude Code

Edit ~/.claude/settings.json and set:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:8080/anthropic"
  }
}

If klovys99 listens on another address, use that instead of 127.0.0.1:8080 in either snippet.

Quick API Checks

Anthropic-style request through Klovys99:

curl http://127.0.0.1:8080/anthropic/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 128,
    "messages": [
      {
        "role": "user",
        "content": "Email Alice at alice@example.com"
      }
    ]
  }'

OpenAI Responses-style request through Klovys99:

curl http://127.0.0.1:8080/openai/v1/responses \
  -H "authorization: Bearer $OPENAI_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "gpt-5",
    "input": "Email Alice at alice@example.com"
  }'

Upstream providers receive the same request shape, with sensitive values replaced by pseudonym tokens such as [EMAIL_1].

How It Works

Klovys99 reads each incoming JSON request body, anonymizes supported prompt content, then forwards the modified request to the configured upstream.

The proxy anonymizes:

every <session>...</session> block found anywhere in a JSON request body;
text content in prompts, system messages, <system-reminder> blocks, text file context, and tool results;
text document sources where source.type is text.

Structural metadata such as model names, roles, content block types, tool IDs, tool names, media types, cache-control values, and base64 document data is left unchanged so the upstream request shape remains valid.

For a single proxy process, repeated values are mapped to stable tokens. For example, the same email address is replaced by the same [EMAIL_N] token across requests handled by that process.

When matches overlap, the detector with the highest priority wins. If priorities are equal, the longest match wins.

Dashboard

While the proxy is running, a built-in dashboard is available at http://127.0.0.1:8080/dashboard. It shows live anonymization stats and links to a test tool at /dashboard/test-tool for previewing how a prompt is anonymized without sending it upstream.

AI Workspace

The AI workspace is a separate chat UI, at frontend/ai-workspace, for talking to a model directly from the anonymized side of the boundary instead of through a coding client. Start it with npm run ui:ai-workspace (see Get Started) while the proxy is running, then open http://127.0.0.1:3001 or use the AI workspace link in the dashboard header.

How it works:

Your prompt is anonymized through the same detectors as the proxy, and the anonymized preview is shown before you send anything.
Only the anonymized prompt is sent to the provider you pick, directly with the API key you configure, not through the /anthropic or /openai proxy routes.
Supported providers are Claude, Gemini, OpenAI, and Mistral, configured with a per-provider API key from the Settings panel in the UI.
Conversations are persisted locally by the Go backend, in plain JSON, under the OS user config directory (klovys99/ai-workspace, override with KLOVYS99_AI_WORKSPACE_DIR).
Saved API keys are only written to disk (encrypted) when KLOVYS99_AI_WORKSPACE_KEY is set. Without it, Settings > Save is a no-op and a key you type only lasts for the current browser tab, resent with each request.
If you want to use the AI workspace UI with persistent provider accounts from the Compte IA / AI accounts panel, set KLOVYS99_AI_WORKSPACE_KEY before starting the backend. A simple way to generate one is:
```
openssl rand -hex 32
```
Then export it, for example:
```
export KLOVYS99_AI_WORKSPACE_KEY="$(openssl rand -hex 32)"
```

The frontend dev server proxies /api and /dashboard to the proxy on 127.0.0.1:8080, and the backend only accepts cross-origin requests from http://127.0.0.1:3001 / http://localhost:3001.

Configuration

Klovys99 runtime is configured with environment variables.

Variable	Default	Description
`KLOVIS_ADDR`	`127.0.0.1:8080`	Listen address for the local proxy.
`KLOVIS_TARGET_URL`	`https://api.anthropic.com`	Upstream used by legacy unprefixed routes such as `/v1/messages`.
`KLOVIS_ANTHROPIC_TARGET_URL`	`https://api.anthropic.com`	Upstream used by `/anthropic/...` routes.
`KLOVIS_OPENAI_TARGET_URL`	`https://api.openai.com`	Upstream used by `/openai/...` routes.
`KLOVIS_PROXY_DEBUG`	`false`	Enables additional sanitized diagnostic logging. Raw traffic is never logged.
`KLOVIS_LOG_PII_FINDINGS`	`false`	Deprecated and ignored for privacy; raw findings are never logged.
`KLOVIS_LOG_TO_FILE`	`false`	Writes logs to `proxy.log` instead of stdout when set to `true`.
`KLOVYS99_AI_WORKSPACE_DIR`	OS user config dir + `klovys99/ai-workspace`	Storage directory for AI workspace conversations and credentials.
`KLOVYS99_AI_WORKSPACE_KEY`	unset	Encryption secret used to persist AI workspace provider API keys to disk. Required if you want the AI workspace UI to keep configured provider accounts across reloads/restarts. Generate one with `openssl rand -hex 32`. Without it, saved keys are not written to disk.

Contextual GLiNER protection modes

The Go binary keeps GLiNER off unless you set KLOVIS_GLINER_MODE=full yourself. Two modes are available and explicit:

full: all configured contextual labels.
off: regex-only mode without contextual GLiNER analysis.

The default pinned model identity is:

model: urchade/gliner_multi_pii-v1
revision: 1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d

./scripts/dev-start.sh builds the sidecar image, installs this pinned model under ~/.klovys99/gliner on first run, starts the sidecar on 127.0.0.1:8091, and then launches the Go proxy in full mode — see Get Started for the equivalent manual docker steps. Skip the sidecar entirely and run go run ./cmd/klovys99 on its own to stay in regex-only (off) mode.

The full mode requests:

person name
organization
location
employer
school or educational institution
medical provider or healthcare institution
street address

A sample direct sidecar latency benchmark is available in docs/benchmarks/gliner-benchmark.md.

Variable	Default	Description
`KLOVIS_GLINER_MODE`	`off`	Explicit contextual mode for the Go binary: `full` or `off`.
`KLOVIS_GLINER_ENABLED`	deprecated	Legacy bool compatibility shim. Prefer `KLOVIS_GLINER_MODE`.
`KLOVIS_GLINER_URL`	`http://127.0.0.1:8091`	Loopback sidecar URL; non-loopback URLs are rejected.
`KLOVIS_GLINER_MODEL`	`urchade/gliner_multi_pii-v1`	Exact model identifier.
`KLOVIS_GLINER_MODEL_REVISION`	`1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d`	Immutable revision/digest.
`KLOVIS_GLINER_TIMEOUT`	`5s`	Per-batch deadline.
`KLOVIS_GLINER_THRESHOLD`	`0.50`	Global confidence threshold.
`KLOVIS_GLINER_LABEL_THRESHOLDS`	`{}`	JSON object overriding thresholds for fixed labels.
`KLOVIS_GLINER_MAX_CONCURRENCY`	`2`	Maximum concurrent inference calls.
`KLOVIS_GLINER_MAX_BATCH_CHARS`	`32768`	Maximum Unicode characters per request batch.
`KLOVIS_GLINER_FAILURE_POLICY`	`fail-closed`	Only supported policy in V1.
`KLOVIS_GLINER_DATA_DIR`	`~/.klovys99/gliner`	Sidecar model directory used by `scripts/dev-start.sh`.

When full is active, a timeout, unavailable sidecar, saturated queue, malformed span, or model identity mismatch returns 503 and makes zero upstream calls. /healthz reports Go liveness, /readyz includes contextual readiness, and /api/status exposes sanitized metadata including the active GLiNER mode.

Boolean variables accept only true or false.

Logs

Klovys99 writes structured application logs to stdout by default. To write logs to proxy.log instead, enable file logging:

KLOVIS_LOG_TO_FILE=true go run ./cmd/klovys99

Debug and file logging remain available, but request bodies, detected values, token mappings, credentials, and model inputs are never logged. The historical KLOVIS_LOG_PII_FINDINGS setting is ignored.

Detectors

Klovys99 combines built-in detectors with external rules loaded at startup. External rule payloads are cached for 24 hours in the user cache directory under klovys99/external-rules.

Category	Source	Priority	Description
`EMAIL`	Built-in / Presidio	1000 / 600	Email addresses, normalized in lowercase for stable tokens.
`NIR`	Built-in	1000	French social security numbers, including spaced formats and Corsica departments `2A` and `2B`.
`IBAN`	Built-in / Presidio	1000 / 600	IBAN-like account identifiers, normalized by removing separators.
`IP`	Built-in / Presidio	900 / 600	IPv4 and IPv6 addresses.
`CREDIT_CARD`	Built-in / Presidio	900 / 600	Credit card-like digit sequences.
`MAC_ADDRESS`	Built-in / Presidio	900 / 600	MAC addresses with `:` or `-` separators.
`PHONE`	Built-in	700	French and common international phone numbers.
`DATE`	Built-in / Presidio	600 / external	Conservatively labelled birth dates and supported contextual dates.
`BLOOD_TYPE`	Built-in	600	Contextual blood groups such as `Groupe sanguin O+`.
`SECRET`	Gitleaks	600	Secrets loaded dynamically from the official Gitleaks config.
`CRYPTO`	Presidio	600	Cryptocurrency wallet identifiers loaded from supported Presidio recognizers.
`ADDRESS`	Built-in	900 / 700	French postal addresses and labelled addresses.
`NAME`	Built-in	900	Contextual names following strong French or English cues and form labels.
`NUMERIC_ID`	Built-in	100	Generic long numeric IDs.
`REFERENCE_ID`	Built-in	100	Labelled alphanumeric references requiring letters and digits.

Claude Code Notes

When Claude Code uses a non-first-party ANTHROPIC_BASE_URL, some Claude features behave differently upstream. In practice:

Remote Control is disabled by Claude Code when the base URL does not point to api.anthropic.com.
Tool search behavior changes when routing through a proxy. If you need deferred tool references, set ENABLE_TOOL_SEARCH=true in your Claude environment because Klovys99 forwards tool_reference blocks unchanged.

Development

Clone the repository and download Go dependencies:

git clone https://github.com/Korbicorp/klovys99.git
cd klovys99
go mod download

Tagged releases build one binary per supported OS and architecture in GitHub Actions. If NPM_TOKEN_KLOVYS is configured in repository secrets, the same tag workflow also publishes the npm package after uploading the release assets — npm installation from that package is currently broken, see Installation.

Run the Go test suite:

go test ./...

The npm packaging code under npm/ has its own unit tests (node --test npm/test/*.test.js) if you're working on that code specifically, but they don't exercise a real install and aren't required to run the proxy.

Run the proxy locally:

./scripts/dev-start.sh

The AI workspace frontend lives in frontend/ai-workspace as its own Node project (installed separately from npm install at the repository root). Install its dependencies once, then use the root scripts to run or build it:

npm --prefix frontend/ai-workspace install
npm run ui:ai-workspace       # dev server on http://127.0.0.1:3001
npm run ui:ai-workspace:build # production build into frontend/ai-workspace/dist

Format Go code before submitting changes:

gofmt -w ./cmd ./internal

Security Notes

Klovys99 reduces the amount of sensitive data sent upstream, but it is not a formal data-loss-prevention guarantee. Review detector coverage for your own threat model before using it with production data.

External Gitleaks and Presidio rules are loaded from their upstream repositories by default. Cached copies are reused for 24 hours and stale cache entries may be used as a fallback if a refresh fails.

Contributing

Issues and pull requests are welcome. For code changes, please include focused tests that cover the behavior being changed.

Useful checks before opening a pull request:

go test ./...
gofmt -w ./cmd ./internal

Name		Name	Last commit message	Last commit date
Latest commit History 69 Commits
.github/workflows		.github/workflows
.vscode		.vscode
cmd/klovys99		cmd/klovys99
docs		docs
frontend/ai-workspace		frontend/ai-workspace
internal		internal
npm		npm
scripts		scripts
sidecar/gliner		sidecar/gliner
test		test
.env.example		.env.example
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
claude.webp		claude.webp
go.mod		go.mod
go.sum		go.sum
klovys99.png		klovys99.png
klovys_tokens.sqlite		klovys_tokens.sqlite
package-lock.json		package-lock.json
package.json		package.json
schema.png		schema.png

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

klovys99

Architecture At A Glance

Get Started

Features

Requirements

Installation

Client Configuration

Codex

Claude Code

Quick API Checks

How It Works

Dashboard

AI Workspace

Configuration

Contextual GLiNER protection modes

Logs

Detectors

Claude Code Notes

Development

Security Notes

Contributing

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

klovys99

Architecture At A Glance

Get Started

Features

Requirements

Installation

Client Configuration

Codex

Claude Code

Quick API Checks

How It Works

Dashboard

AI Workspace

Configuration

Contextual GLiNER protection modes

Logs

Detectors

Claude Code Notes

Development

Security Notes

Contributing

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages