Strac MCP DLP¶

Open-source MCP server (stdio, or Streamable HTTP as a service) that puts a DLP check in front of the model. Agents are now wired into inboxes, ticketing systems, CRMs and file stores - every tool call is a chance for an SSN, a card number, a patient record or an AWS key to be pulled into a prompt and from there into a model provider's logs. Strac's server detects first and redacts before the agent reasons over the text, using the Strac DLP API for classification. Five tools, thin client, MIT-licensed repository.

Server type: Local (stdio via pip install strac-mcp-dlp; --transport streamable-http for services)
Auth: Strac API key (request at strac.io/mcp-integrations)
Tools: 5 (redact_text, detect_sensitive_data, detect_file, redact_file, detokenize)
Pricing: Free tier; classification runs server-side at Strac
Category: Security / Data Protection
Built by: Strac (strac.io)

Why This Matters for Operators¶

Filtering sensitive data AFTER the model has seen it is too late - it is already in a provider's logs, a vector store or a support ticket. This server moves the check in front of the model: detect, redact, then let the agent work on text that no longer carries the values. The classification catalog covers personal, health and payment data plus secrets - AWS access and secret keys, GitHub, GitLab and Slack tokens, GCP credentials, Azure keys, private keys, and JDBC and MongoDB connection strings.

Redaction that does not leak. By default the tools return the TYPES and POSITIONS of what they found, not the values - handing the raw values back would undo the redaction. include_matched_text=true is an explicit opt-in. redact_file never modifies the original: it writes a copy, refuses an output path that resolves to the source, and will not overwrite an existing file without permission.

Tools & Capabilities¶

Tool	What it does
`redact_text`	Redact PII/PHI/PCI/secrets from text; modes REDACTED, BLANK, MASK_SEVEN_X, TOKEN_LINK_PLAINTEXT
`detect_sensitive_data`	Report which sensitive types are present, unchanged
`detect_file`	OCR + classify a local image, PDF, scan, config or source file
`redact_file`	Write a redacted copy to disk; original untouched
`detokenize`	Resolve Strac vault tokens back to values for authorized, IP-allowlisted callers

Files under 4 MB are classified inline without storage; larger files go to your Strac document vault. On an unparseable API response the tool errors - it never reports a clean scan it could not verify, because a silent false negative is the one failure a DLP tool cannot have.

Installation¶

pip install strac-mcp-dlp

Configure it as a stdio MCP server in Claude Desktop, Claude Code, Cursor or VS Code. Run it as a service with --transport streamable-http if you need a shared endpoint. Bring the Strac API key from strac.io/mcp-integrations; every call is authenticated, no anonymous mode.

Configuration¶

The key rides in the server environment. redact_text takes a redact_field_mode (default REDACTED) and the opt-in include_matched_text. detect_file/redact_file take local paths; nothing is classified locally - all detection runs server-side over HTTPS.

Business Relevance¶

Compliance-adjacent operations (finance, healthcare-adjacent SaaS, agencies handling client PII) get a cheap first-line control: agents that process support tickets, contracts, invoices or HR files can be pointed at Strac so sensitive values never enter the prompt. It complements the SaaS-side Strac product, which discovers and remediates data across 60+ integrations under policy.

Integration with CorpusIQ¶

CorpusIQ agents read sensitive business data (invoices, transactions, payroll) across 40+ connectors. Pairing a DLP gate like Strac means that when an agent composes a summary or a report, the sensitive fields are redacted or tokenized at the tool boundary before they enter the model context - preserving the analysis while keeping PII out of model logs and downstream artifacts.

Limitations¶

All classification runs in Strac's cloud - no local detection models; data in transit is HTTPS
detokenize requires an IP-allowlisted server-to-server key in live mode
Redaction fidelity depends on Strac's classifiers; positions, not values, are the default return