ClaimCheck
Local-first secret shield for AI coding agents: tickets passwords, API keys, tokens and client PII before they reach the LLM — reversible via a local encrypted vault. MCP tools + CLI, zero network calls.
ClaimCheck — local PII pseudonymization
ClaimCheck replaces sensitive values with stable, reversible placeholders
([FEIN_1], [PASSWORD_2], [PERSON_1]...) and keeps the mapping in a local
encrypted vault. Nothing leaves the machine — no account, no API key, no
network calls.
Why: the debug-leak problem
Every time an agent debugs a live project it reads .env, appsettings.json,
logs and scripts — and those secrets ride into the LLM provider's stored
context. ClaimCheck tickets them before they leave: the model works over
tickets, your secrets stay local, and files the agent writes come back with
the real values restored.
When to use (always for client-adjacent data)
- Before quoting any file's contents into chat, tickets, docs, emails, or commits when it may contain client data (FEIN, SSN, account numbers, names, credentials)
- Before sharing sample data — share the sanitized sample instead
- To CHECK a file for PII without reading it:
sanitize-file <path>(no --content)
Commands
# text from stdin → JSON {label, summary, findings, sanitized}
echo "..." | dotnet bin/Release/net9.0/ClaimCheck.dll sanitize-text
# text as argument
dotnet bin/Release/net9.0/ClaimCheck.dll sanitize-text "some text"
# file → summary only (types/counts/lines, no content) or with sanitized content
dotnet bin/Release/net9.0/ClaimCheck.dll sanitize-file /path/to/file --content
# restore placeholders → real values (local vault only)
dotnet bin/Release/net9.0/ClaimCheck.dll restore <file|->
Recognized (regex tier, always on): FEIN, SSN, email, phone, cards (Luhn), IBAN, AWS/GitHub/Slack tokens, API keys, JWTs, connection-string passwords, PEM private-key blocks, base64/hex blobs, .env/config label anchoring (KEY/SECRET/TOKEN/PASSWORD/COOKIE labels), provider prefixes (github_pat_, whsec_, sess_, npm_, glpat_).
Optional tiers:
- Presidio (local container,
SANITIZE_PRESIDIO_URL): free-text NER + crypto wallets, country formats. Skipped above 50K chars (regex tiers still full). - Known-value re-ticketing: once a value is in the vault it can never reach the model again in any form — re-reads, echoes, and tool outputs all re-ticket.
Rules
- Share ONLY the
sanitizedfield — findings show placeholders, never values. - Never paste raw FEIN/SSN/card/credential values into chat, logs, or commits.
- The vault lives in
~/.sanitize-mcp/vaults(encrypted at rest on Windows via DPAPI; AES-GCM withSANITIZE_VAULT_PASSWORDelsewhere). The machine is the trust boundary. - Not 100% — verify sanitized output before high-stakes sharing (DISCLAIMER.md).
On-wire proxy (scrub any OpenAI-compatible client)
The proxy sits between your client and the LLM provider: every request is scrubbed, every response restored. Works with Cline, Cursor, Codex CLI — any client that takes a custom Base URL.
Windows (one-time): install the GitHub release — install.ps1 registers a
logon task that runs the tray, which hosts the proxy on :8766.
Any OS (manual): run the binary as the proxy:
ClaimCheck --proxy --upstream https://api.z.ai/api/coding/paas/v4/ --port 8766
Point your client's Base URL at the proxy — it mirrors the upstream path:
http://127.0.0.1:8766/api/coding/paas/v4 (z.ai)
http://127.0.0.1:8766/v1 (OpenAI/Anthropic/OpenRouter)
Keep your real API key in the client — auth passes through untouched. Optional tiers on the proxy: SANITIZE_PRESIDIO_URL (local presidio container) + SANITIZE_NER=1 (local NER model). Full story: https://claimcheck.saasclaw.ai
- Downloads
- Security audit
- Review
- Last updated
- 3w ago
- Current version
- v1.1.3
- License
- MIT-0
