🛡️ SecureMCP

Local English/Korean Masking · Subscription Gateway · Trusted Restoration

영어·한국어·코드를 로컬에서 마스킹하고, 도구 실행과 사용자 화면에서 원문을 복원하는 개인정보 보호 도구

English · 한국어 안내 · Supported Versions · Quick Start · Features · Architecture · Documentation

English Overview

SecureMCP masks English/Korean text and code locally, with session-based restoration.

Version 0.5 adds an experimental local subscription gateway for Claude Code and Codex. Run secure-mcp gateway install once, then use claude / codex normally. Requests are masked automatically; responses and local tool arguments are restored automatically. Existing CLI subscription OAuth is forwarded to the original subscription service; API keys are refused. No agent hooks are installed. See subscription gateway for setup, automatic startup, tested versions and the exact protection boundary.

Legacy agent hooks remain optional utilities. They cover selected tool text rather than complete requests; local integration documents their limits.

Mask confidential input before sending it to a provider. Restore responses in your trusted local application and show them to the user there. Model-invoked MCP tool arguments are already visible to the provider; adding this server to Claude Desktop or Cursor does not automatically intercept or protect prompts. Restored originals must not be sent back into the model context. This is heuristic masking, not a proof of anonymity, cryptographic zero knowledge, or regulatory compliance.

Related MCP server: ai-security-gateway-mcp

Supported Versions

The CI matrix covers 15 Python/OS combinations. Code-language support describes lexer modes, not compatibility with every language release or compiler. The gateway is experimental. Native fixture tests verify routing/masking/restoration. Codex passed a live subscription smoke test. Claude's attribution masking bug is fixed; live requests now return HTTP 200, but a Sonnet safety-filter refusal prevents successful answer/restoration verification. The earlier 429 was not subscription exhaustion. This does not certify all-traffic privacy.

Quick Start

Automatic subscription integration

uv tool install .
secure-mcp gateway install
secure-mcp gateway status

Restart Claude Code/Codex and use their normal commands. Setup preserves login caches and registers a hidden user-login startup task. It changes user-wide provider settings; no agent hooks are added. Subscription limits still apply. This is text/code inference protection; images, remote attachments and unsupported payloads are blocked. See gateway guide.

Legacy optional agent hooks

uv tool install .
secure-mcp init --agent claude
secure-mcp doctor

Restart Claude Code after installation. Hooks default to the current project; use --global for user-wide registration. See local integration for supported tool fields, installation/removal and limitations.

For Codex:

secure-mcp init --agent codex
secure-mcp doctor --agent codex

Restart Codex and review/trust the installed definitions in /hooks. Project hooks require a trusted project. Codex masks tool results and restores Bash/apply_patch inputs; automatic screen restoration is unavailable. Use secure-mcp restore --agent codex --session-id <id> with masked text on UTF-8 stdin for local display. See Codex integration.

Development installation and demo

uv venv
uv pip install -e ".[dev]"
uv run python -m secure_mcp demo

Trusted local Python API

from secure_mcp import LocalPrivacyClient

# Replace echo with your provider adapter. Only its argument may leave the host.
def provider(masked_payload: str) -> str:
    return masked_payload

with LocalPrivacyClient() as client:
    result = client.request(
        "Patient John Doe received 50mg.", provider,
        sensitive_terms={"John Doe"},
    )
    print(result.unmasked_text)  # Local display only

Features

Masking policies

English, Korean and mixed input are handled per word. Unicode names remain atomic. Korean particle splitting uses known stems and conservative rules; unknown ambiguous words stay whole. It is not a full morphological analyzer or universal person-name detector. Supply sensitive_terms for domain names, lowercase names, ambiguous names, and custom secrets. custom_preserve / CLI --preserve deliberately exempts selected words; recognized sensitive spans still take precedence. Unknown secret formats and context-dependent names may evade entities_only; inspect the payload or use broader masking. A high masking ratio does not prove absence of sensitive data.

Surrogate strategies

Pseudowords use explicit delimiters to avoid collisions with real words and adjacent tokens. Allocation is stable within a session; random strategies intentionally differ between sessions. Keep all surrogate spelling intact. Altered/unknown placeholder candidates are reported in unmatched_surrogates; strict=True rejects them. Free-form deletion or invention by a model cannot always be detected or reconstructed.

Korean restoration

Exact mask/unmask roundtrips keep written particles. For newly generated Korean responses, opt into normalize_particles=True (CLI --normalize-particles) to choose 이/가, 은/는, 을/를, 과/와 and 으로/로 from a restored Hangul stem. Pronunciation of foreign names is not guessed.

Code-aware review representations

Code languages: python, javascript, typescript, go, rust, java, c, cpp, sql. Use explicit language in mask_code (CLI --code-language) for ambiguous snippets. Auto detection is best effort. Keywords are language-specific; builtins and shadowed builtin names are masked. Unicode identifiers, SQL comments/doubled quotes, backticks, Python f-strings and escaped newlines, C++/Rust raw strings and numeric suffixes/separators are covered by tests. Code allocations use ASCII names (smcp_ID_n and smcp_LIT_n inside literals) and random native integer constants for numbers (reserved 732846 prefix plus 12 digits), independently of the text strategy, so byte literals remain valid too. Python output is syntax-checked in tests. Output is an opaque review representation: it need not execute, type-check, retain numeric types/values, or preserve f-string/template interpolation behavior. This small lexer is not a complete parser for every version of every supported language.

System Architecture

flowchart LR
    Input[Trusted local input] --> Mask[Masking engine]
    Mask -->|Masked payload| Provider[AI provider callback]
    Provider -->|Surrogate response| Restore[Local restoration]
    Restore --> Display[User display]
    Vault[Local session mappings] --- Mask
    Vault --- Restore

The subscription gateway masks configured inference requests before forwarding them and restores responses before the CLI consumes them. The callback path masks its provider argument. Legacy agent hooks use a separate, partial tool-text path; they do not intercept all outgoing context. See the architecture and local integration documents.

CLI and Sessions

python -m secure_mcp mask "Alice from Google" --session-id example --session-file example.enc --json-output
python -m secure_mcp unmask "[ENT_1] from [ENT_2]" --session-id example --session-file example.enc --strict
python -m secure_mcp mask "john doe" --mode entities_only --sensitive-term "john doe"

Use the same hidden password, or SECURE_MCP_SESSION_PASSWORD. CLI files use authenticated Fernet encryption, random salt, PBKDF2-HMAC-SHA256 with 600,000 iterations, atomic writes and a lock over the complete CLI read/modify/write operation. A concurrent writer fails clearly. A crash may leave example.enc.lock; remove it only after confirming no writer is running. Files are opt-in; without --session-file, mappings only survive in the current process. Payload schema v2 stores allocations/counters without regenerating mappings. Legacy v1 files remain readable; unknown schema versions fail explicitly. Delete session files after use. Legacy bare pseudoword allocations retain their old ambiguity; start a new session to use the safe delimited representation for every allocation.

Session strategies are immutable; mode records the last-used policy. Creating a duplicate ID fails without changing its TTL or mappings. TTL must be positive and at most one day. The vault sweeps idle expired sessions at most one cleanup interval later (default one second), and checks expiration on access. Clear waits for active operations and invalidates retained references. Library users must use session.operation() for custom mutations; built-in engine operations use the same lock. vault.close() stops cleanup and releases all mapping references. Python cannot promise byte-level memory zeroization. Standalone PrivacySession owners manage lifetime themselves; use SessionVault for scheduled expiry.

masked_ratio counts masked token occurrences divided by non-whitespace/non-punctuation occurrences. Deprecated privacy_entropy_score is an alias, not information entropy. restored_occurrences counts replacements; restored_unique_tokens counts distinct allocations. restored_tokens_count remains a compatibility alias for replacement occurrences.

MCP Tools Reference

python -m secure_mcp serve supports local stdio only. HTTP/SSE transports are disabled in the CLI; directly exposing the Python server over a network requires host-provided authentication and isolation.

Local Python unmask_text / unmask_code and CLI unmask are not registered as MCP tools, so a model cannot enumerate mappings via restoration. Resources: privacy://policies, privacy://status. Example desktop configurations are utility setups, not privacy proxies.

Migration from 0.1

Version 0.2 changes pseudoword delimiters, code identifier/number representations, mapping keys (mapping_key(original, token_type, code=...)), sentence-initial entity classification, keyword preservation, and the MCP restoration boundary. Direct store readers should use mapping values or mapping_key. Validate explicitly requested strategies; omitted strategies follow the generator.

한국어 안내 (Korean Overview)

0.5에서는 Claude Code·Codex의 기존 구독 로그인을 유지하는 로컬 게이트웨이를 제공합니다. uv tool install . 설치 후 secure-mcp gateway install을 한 번 실행하고 두 CLI를 재시작하세요. 그 뒤에는 평소처럼 claude / codex를 사용하면 됩니다. 직접 입력·텍스트 코드·도구 결과를 자동으로 마스킹하고, 답변과 로컬 실행 인자는 자동 복원합니다. 함수명·변수명·문자열·숫자·주석을 보존하는 예외는 추가하지 않았습니다. 입력에서 지칭한 함수와 코드의 함수는 같은 치환표를 사용합니다.

기존 로그인 파일은 읽거나 복사하지 않습니다. CLI가 관리하는 OAuth 헤더와 갱신 흐름을 사용하고, API 키 요청은 거부하므로 유료 API로 자동 전환하지 않습니다. 새 Hook을 등록하지 않으며, 운영체제 사용자 로그인 시 백그라운드로 자동 시작하도록 설정합니다. 기존 Hook 방식은 선택 기능으로 남겨 두었습니다. 설치·해제 및 보호 범위를 확인하세요.

Claude Code 2.1.287·Codex 0.160.0의 실제 CLI를 모의 OAuth/로컬 서버로 검증했습니다. Codex는 실제 구독 요청·자동 복원까지 확인했습니다. Claude는 식별 블록 처리 오류를 수정해 실제 구독 HTTP 200을 확인했지만, Sonnet 안전 필터 거절로 정상 답변·복원 검증이 남았습니다. 기존 429는 구독 한도 소진이 아니었습니다. 이미지·원격 첨부·미지원 요청은 차단하고, 게이트웨이 밖의 도구 네트워크·텔레메트리까지 보호한다고 주장하지 않습니다.

SecureMCP는 영어·한국어·혼합 문장을 로컬에서 마스킹하고 복원합니다. 클라우드에 요청하기 전에 LocalPrivacyClient 또는 로컬 API/CLI로 원문을 가리고, 응답은 로컬에서만 복원하세요. 모델이 호출하는 MCP 도구의 인자는 이미 제공자에게 전달되어 있으므로, Claude Desktop/Cursor에 서버를 추가하는 것만으로 개인정보가 보호되지는 않습니다. 복원한 원문을 모델 컨텍스트에 다시 넣으면 안 됩니다. MCP 복원 도구는 제거했으며, 기본 CLI 서버는 로컬 stdio만 지원합니다.

영어 대문자 이름과 Unicode 이름을 하나의 토큰으로 처리합니다. 한국어 이름과 일반 명사가 구분되지 않는 entities_only에서는 비기능어를 보수적으로 가립니다. 조사 분리는 알려진 어간과 보수적인 규칙을 사용하며, 모호한 미등록 단어는 통째로 가립니다. 모든 고유명사·비밀값 또는 한국어 형태소를 완벽히 인식하는 도구는 아닙니다. sensitive_terms / --sensitive-term으로 특정 용어를 지정하고, 알 수 없는 민감정보에는 더 넓은 마스킹 정책을 사용하세요.

정확한 왕복 복원에서는 입력 조사를 그대로 유지합니다. 모델이 새로 만든 문장에는 normalize_particles=True / --normalize-particles를 선택하여 받침에 맞는 조사를 보정할 수 있습니다. 영어 이름·약어의 발음은 추측하지 않습니다. --strict는 탐지한 변형·미등록 플레이스홀더의 복원을 거부하지만, 모델이 완전히 삭제한 내용을 재구성하지는 못합니다.

세션 전략은 생성 후 고정되며, 중복 ID 생성은 오류입니다. 만료 세션은 주기적으로 정리하고, 삭제와 진행 중 작업을 잠금으로 조율합니다. 별도 CLI 프로세스 간 복원에는 동일한 암호화 --session-file이 필요합니다. 마스킹 비율은 통계이며 기밀성·익명성·규제 준수의 증명이 아닙니다.

Development

python -m ruff check src tests examples
python -m mypy src
python -m pytest -W error --cov=secure_mcp --cov-report=term-missing --cov-fail-under=85

Documentation

• Subscription Gateway / 구독 로그인 자동 연동

• Legacy Local Claude Code & Codex Integration / 기존 Hook 안내

• System Architecture

• Issue Resolution and Review Notes

• Contributing Guidelines

License

Licensed under Apache License 2.0. See LICENSE for details.