Filtir

Filtir is an asynchronous reverse proxy written in Rust, designed to intercept traffic from LLM clients, detect and redact personally identifiable information (PII) using regular expressions, and then forward the cleaned request body to the actual API. A local web interface (dashboard) allows for real-time configuration adjustments without needing to recompile the executable.

1. Project Operation

Architecture

• Two HTTP services run within the same binary on two different ports:

  • Proxy port (input) The endpoint for clients (CLI, editors, scripts) that previously called the provider's API directly. Default: 3100.
  • Dashboard port Configuration interface + REST API. Default: 3200.

• The application state (AppConfig statistics, logs) is shared via an Arc and locks, allowing the dashboard and proxy to read/write the same logic (PII rules, output URL, etc) .

Proxy Request Flow

1. The client sends an HTTP request to the input port (using the same path and query as the target API, e.g., POST /v1/chat/completions).
2. The body is read and interpreted as JSON: all strings within the JSON tree are processed through the active PII rules. If the body is not valid JSON, it is treated as plain text.
3. Each active rule is a precompiled regex; matches are replaced with the replacement string (e.g., [EMAIL REDACTED]).
4. The modified body is forwarded via reqwest to the Output URL + the original path (e.g., https://api.openai.com/v1/chat/completions).
5. The response from the backend API is returned to the client. The X-Filtir-Redacted header indicates the number of detected matches.

What Filtir Does Not Do

• No persistence on disk: a restart resets the configuration to default.
• Redaction applies to the body of requests, not a global network audit of the system.

2. Using the Dashboard

1. Start the executable (see Section 3), then open http://localhost:3200 in a browser.
2. Statistics Cards: Processed requests, total PII matches, active rules. Counters refresh automatically.
3. Configuration: Modify the proxy input port and output URL, then save. Changing the input port triggers a graceful shutdown of the current listener and a restart on the new port (the dashboard does not restart).
4. PII Rules: Each rule has a name, a regex pattern, replacement text, and an on/off switch. You can add, modify, or delete them on the fly.
5. Redaction Test: Paste free text and click "Test" to see the redacted result and the count of detected PII without going through the proxy.
6. Request Log: The last 100 requests seen by the proxy (time, method, path, PII count, HTTP code).

API REST du dashboard

Méthode	Chemin	Rôle
GET	/	Serves the HTML interface.
GET	/api/config	Returns the current configuration (JSON).
POST	/api/config	Updates the config; restarts the proxy if the input port changes.
GET	/api/stats	Global counters (requests, PII).
GET	/api/logs	Last 100 requests.
POST	/api/test-redaction	Body: {"text":"..."}. Response: original, redacted, count.

3. Installation

Prerequisites

• Rust (2021 edition, latest stable): install viarustup.rs.
• No additional system dependencies (no OpenSSL): the HTTP client uses rustls, which is built into the binary.

Compile

git clone https://github.com/Improba/Filtir filtir
cd filtir
cargo build --release

The binary is available in target/release/filtir (or target\release\filtir.exe on Windows).

Run

./target/release/filtir
# or in development mode:
cargo run

The default log level is info. For more detail:

RUST_LOG=filtir=debug ./target/release/filtir

Stopping : Ctrl+C (graceful shutdown of both servers).

Quick Verification

# Is the dashboard responding?
curl -s http://localhost:3200/api/stats

# Test redaction directly from the terminal:
curl -s -X POST http://localhost:3200/api/test-redaction \
  -H "Content-Type: application/json" \
  -d '{"text": "Call John Doe at 06 12 34 56 78, email: john@example.com"}'

4. Configuring Claude Code and Other Tools

The principle: point the client to http://127.0.0.1:3100 (Filtir's proxy port) instead of the remote API, and specify the real API URL in the Output URL field of the dashboard.

---

4.1. Claude Code (CLI)

Claude Code natively supports ANTHROPIC_BASE_URL. Since calls are made directly from your machine, 127.0.0.1 works without any network exposure.

Recommended Method : ~/.claude/settings.json

Create or modify ~/.claude/settings.json :

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:3100",
    "ANTHROPIC_API_KEY": "sk-ant-your-real-key"
  }
}

This file is loaded every time claude is launched. The real API key is transmitted via Filtir to Anthropic in the header.

One-off Method: Environment Variable

# For the entire shell session
export ANTHROPIC_BASE_URL="http://127.0.0.1:3100"
claude

# For a single command
ANTHROPIC_BASE_URL="http://127.0.0.1:3100" claude -p "Bonjour"

# To disable the proxy for one command (without touching exports)
env -u ANTHROPIC_BASE_URL claude

Filtir Dashboard Setting

In the dashboard (http://localhost:3200), Output URL :

https://api.anthropic.com

If non-essential traffic errors appear

Claude Code performs startup requests (autoupdater, telemetry) that may fail via a strict proxy. Add to settings.json :

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:3100",
    "ANTHROPIC_API_KEY": "sk-ant-your-real-key",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Verification

claude -p "My professional email is alice@acme-energy.fr and my IBAN is FR76 3000 1234 5678 9012 345"
# Claude's response should arrive normally.
curl -s http://localhost:3200/api/stats
# "pii_blocked" should have increased by 2.

---

4.2. Cursor

Important Limitation: Cursor assembles its prompts via its own backend servers, then calls the provider API from those servers. When you enter an "Override Base URL" in Cursor, it is Cursor's backend—not your machine—that contacts that URL. http://127.0.0.1 would therefore point to Cursor's servers, not your local Filtir.

Cursor's "Override OpenAI Base URL" option is only usable with Filtir if your proxy is exposed on a network address accessible from the outside (see 4.2.2). For purely local use without network exposure, extensions that make calls directly from the client are better suited (Continue.dev, Cline — see 4.3).

4.2.1. Locating the Setting in Cursor

1. Open Cursor settings: Ctrl+, (Linux/Windows) or Cmd+, (macOS).
2. n the sidebar, click on Models.
3. Find the OpenAI API Key section.
4. Enter your OpenAI key, then enable the "Override OpenAI Base URL (when using key)" toggle.
5. In the field that appears, enter your proxy URL (see 4.2.2 below).
6. Click Verify to test the connection.

This setting only affects OpenAI models configured with your own key. Tab-completion uses Cursor's internal models and never passes through this proxy.

4.2.2. Temporarily Exposing Filtir with ngrok (for testing)

# Terminal 1 : launch Filtir
./target/release/filtir

# Terminal 2 : expose the proxy port
ngrok http 3100

ngrok will display a public URL like https://xxxx-xx-xx.ngrok-free.app. Enter this into Cursor's Override OpenAI Base URL field. Cursor adds /v1 and the request paths itself—so do not add /v1 to the provided URL.

In the Filtir dashboard, URL de sortie :

https://api.openai.com

---

4.3. Continue.dev (VS Code / Cursor Extension)

Continue is an open-source extension whose API calls are made directly from your machine, making 127.0.0.1 fully functional without network exposure.

1. Install the Continue extension from the VS Code Marketplace (compatible with Cursor).
2. Open ~/.continue/config.json and add a model pointing to Filtir:

{
  "models": [
    {
      "title": "GPT-4o via Filtir",
      "provider": "openai",
      "model": "gpt-4o",
      "apiKey": "sk-your-openai-key",
      "apiBase": "http://127.0.0.1:3100/v1"
    }
  ]
}

3. In the Filtir dashboard, Output URL : https://api.openai.com.
4. In the Continue panel, select the "GPT-4o via Filtir" model.

---

4.4. Scripts and CLI Tools (Python SDK, curl, etc.)

Any tool that respects standard OpenAI environment variables:

export OPENAI_BASE_URL="http://127.0.0.1:3100/v1"
export OPENAI_API_KEY="sk-your-openai-key"

Example with curl :

curl -s http://127.0.0.1:3100/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "model": "gpt-4o",
    "messages": [{
      "role": "user",
      "content": "Contact John Doe at 06 12 34 56 78, IBAN FR76 3000 1234 5678 9012 345"
    }]
  }'

Example with Python SDK :

import openai  # pip install openai

client = openai.OpenAI(
    base_url="http://127.0.0.1:3100/v1",
    api_key="sk-your-openai-key",
)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "My email is alice@energy.fr, PRM: 12345678901234"}]
)
print(response.choices[0].message.content)

In the Filtir dashboard, Output URL : https://api.openai.com.

---

4.5. Compatibility Summary

Tool	Calls from Local Machine	127.0.0.1 Proxy Functional	Mechanism
Claude Code CLI	Yes	Yes	ANTHROPIC_BASE_URL in ~/.claude/settings.json
Continue.dev (extension)	Yes	Yes	apiBase in ~/.continue/config.json
Cline (VS Code extension)	Yes	Yes	"Base URL" field in extension settings
OpenAI Python / Node SDK	Yes	Yes	base_url parameter or OPENAI_BASE_URL variable
curl / shell scripts	Yes	Yes	Explicit URL in command
Cursor	No (serveurs backend Cursor)	No	Requires a public URL (ngrok, etc.)

---

4.6. Points d'attention

• Local HTTP: Filtir listens on clear HTTP. Do not expose it on a public interface without a firewall.
• TLS to target API: Filtir handles outbound HTTPS (via rustls); no TLS configuration is required on your end.
• Order of Rule Application: PII rules are applied in the order shown in the dashboard. Reorganize them if two patterns might overlap.
• Cursor Tab-completion: Uses Cursor's internal models and never goes through Filtir.
• Validation: Use POST http://localhost:3200/api/test-redaction to test your regex before connecting a client.