PrivacyFence

Human control and policy enforcement for AI access to enterprise data.

PrivacyFence is a local enterprise AI governance layer for macOS. It sits between an MCP-compatible AI assistant and the business systems it can access, so data reads and actions are reviewed, governed, and logged before they are executed.

Instead of granting an AI assistant broad, persistent access and relying on the assistant to use it safely, PrivacyFence applies an independent control point:

• Human approval for sensitive reads and actions
• Policy-based automation for routine, low-risk requests
• PII detection before personal data enters the AI context
• Audit logging for accepted, denied, and automatically approved requests
• Connector-level control across common enterprise systems
• Local credential ownership: credentials remain in the PrivacyFence daemon, not in the AI-facing bridge

PrivacyFence currently runs on macOS and integrates with Claude through MCP. Its governance model is designed around a broader problem: controlling how AI assistants access and act on enterprise information.

---

Why PrivacyFence?

Organizations are rapidly adopting AI assistants, but conventional access-control models were designed for people and applications—not autonomous agents that can independently search, retrieve, combine, and modify information across multiple systems.

Granting an assistant access to Gmail, Drive, Slack, Salesforce, Jira, or other business tools can create a new gap between authorization and intent:

• A user may be authorized to access a record, but may not want that record sent to an AI.
• A connector may technically permit an action, but the action may still require human judgment.
• Static permissions cannot capture the context of a specific request.
• Native AI-client prompts may show technical tool names without enough business context.
• Broad access can reduce accountability unless every decision is traceable.

PrivacyFence introduces a governance layer between the assistant and enterprise systems. It allows AI to remain useful while keeping access decisions visible, contextual, and accountable.

---

Governance principles

PrivacyFence is built around five principles:

1. The AI assistant is not the authorization boundary.
   Policy and approval are enforced independently by PrivacyFence.

2. Humans stay in control of sensitive data and consequential actions.
   Users can inspect the actual content or action before approving it.

3. Routine work should remain efficient.
   Narrow auto-accept rules and temporary approvals reduce repetitive prompts without creating unrestricted access.

4. Every decision should be auditable.
   Accepted, denied, and auto-accepted requests are recorded locally.

5. Governance should enable AI adoption—not block it.
   The objective is safe, practical use of AI across real business workflows.

---

What PrivacyFence governs

PrivacyFence applies policy and review to both directions of an AI workflow.

Data flowing to the AI

Examples:

• Reading an email or email thread
• Opening a Drive file or Google Doc
• Reading spreadsheet values
• Searching Slack or Telegram
• Reading a Jira issue
• Fetching a Salesforce record or report
• Reading a Confluence page or downloading its attachments

Sensitive reads can be shown in full before release to the assistant. The optional PII detection gate scans read content locally and adds an explicit warning when likely personal data is found.

Actions flowing from the AI

Examples:

• Creating a Gmail draft
• Sending a Slack or Telegram message
• Creating or updating calendar events
• Writing to a Drive file, Google Doc, or spreadsheet
• Creating or updating Jira issues
• Creating or updating Confluence pages
• Creating or updating contacts and tasks
• Moving or uploading files

Write dialogs explain the intended operation in business terms and require explicit approval before execution.

---

Human-readable approvals

PrivacyFence translates MCP tool calls into operation-specific review dialogs. Users see the target object, relevant metadata, the full content or action, and the available decision—not just a raw tool name or JSON payload.

PII-aware review

The PII gate runs locally before a read is approved. When likely personal data is detected, PrivacyFence highlights the categories found and requires an additional confirmation. Every review dialog also discloses exactly what the AI will receive, how often the item has come up recently, and the reason Claude gave for the request.

Review before an AI action

Write operations show exactly which object will change and what values will be written, including an informational warning when Claude's own drafted content appears to contain sensitive figures. Selected high-frequency operations can receive a narrowly scoped, in-memory Accept for 5 min approval.

Local administration

The menu bar provides access to:

• PII detection
• Organization configuration
• Auto-accept rules
• Connector authentication and status
• The local audit log

Auto-accept rules and trusted-resource grants are managed from a single searchable window, organized by connector:

---

Policy-based automation

Not every request needs a popup.

PrivacyFence can automatically approve routine, low-risk operations when a narrowly defined rule matches. Rules can use context such as:

• sender domain or Gmail label
• file ownership or approved Drive folder
• spreadsheet and tab
• Slack channel
• calendar ownership and attendee scope
• Jira project
• Confluence space
• Salesforce object type or report
• Telegram chat
• Google Tasks list

PII detection takes precedence over read auto-accept rules. A request that would normally pass silently is returned to human review when likely personal data is found in the content.

Temporary approval is also available for selected repetitive write operations. These approvals are scoped to the same operation and file, held only in memory, and expire automatically.

Scheduled Claude Cowork tasks can run unattended: a privacyfence_check_policy tool lets Claude check ahead of time whether a call would auto-accept or need a human, and an opt-in unattended-session mode denies unmatched requests immediately instead of leaving a popup open for nobody to answer. See Technical Reference.

---

Supported connectors

Connector	Examples of governed capabilities
Gmail	Read messages and threads, download attachments, create drafts and replies (plain text or rich text/HTML), manage labels, archive messages, create and update filters
Google Drive, Docs & Sheets	Read, download, upload, move, and write files; write, partially edit, and format (including highlight) Google Docs; add comments; read, write, and format Sheets ranges; add and rename tabs; insert and delete rows/columns
Google Calendar	Read event details, get/set event visibility, create and update events, create out-of-office entries, set working location
Google Contacts	Read, create, update, and label contacts
Slack	List channels, DMs, and group chats (filterable by participant), read channels and threads, search messages, send messages
Salesforce	Read records, search by name or id, and run reports
Jira	Read, create, update, comment on, and transition issues
Confluence	Read, create, and update pages; list and download page attachments
Google Tasks	Read, create, update, complete, uncomplete, and move tasks
Google Apps Script	Read and write script project source; read the result of a run you triggered yourself (PrivacyFence never runs scripts)
Telegram	Read chats, search messages, and send messages

The detailed tool-by-tool privacy matrix is maintained in Technical Reference.

---

Architecture

PrivacyFence is one persistent macOS daemon, privacyfence-app, which owns credentials, connector clients, policy evaluation, approval dialogs, PII detection, temporary approvals, and the audit log. It exposes a local, token-authenticated /mcp Streamable HTTP endpoint that Claude talks to directly (Claude Code) or through a thin stdio-to-HTTP shim PrivacyFence.mcpb installs (Claude Desktop, which has no built-in way to reach an HTTP MCP endpoint directly — see the refactor plan §12, D11). The shim carries no service credentials and no tool-schema knowledge of its own; it only forwards bytes.

This replaced an earlier design where a small, ephemeral privacyfence-bridge Node process, started by the AI client, forwarded requests over a local TCP loopback socket instead of HTTP — retired once both transports had shipped for a full release cycle each (P5 of the refactor plan above).

Claude Code          Claude Desktop
     │                     │
     │ MCP over HTTP       │ MCP over stdio
     │                     ▼
     │              stdio<->/mcp shim
     │                     │
     │  local, token-authenticated /mcp
     ▼                     │
     └──────────┬──────────┘
                ▼
         privacyfence-app
                │
                ├── policy and auto-accept rules
                ├── PII detection for reads
                ├── human review gate
                ├── temporary approvals
                ├── audit log
                └── enterprise connectors

See Technical Reference for the review model, connector matrix, rule catalogue, installation, configuration, /mcp design, and MCP annotation rationale.

---

Who is PrivacyFence for?

PrivacyFence is relevant to:

• CIOs and CTOs introducing AI assistants into business workflows
• CISOs and information-security teams
• AI governance, privacy, risk, and compliance leaders
• Enterprise IT administrators
• Developers building MCP-enabled workflows
• Organizations assessing AI use under GDPR, the EU AI Act, and internal information-handling policies

PrivacyFence is currently an open-source macOS implementation rather than a certified compliance product. It can support governance and evidence collection, but it does not by itself make an organization compliant with any regulation.

---

Why existing approaches fall short

Existing approach	PrivacyFence
Block AI access entirely	Enable AI through governed access
Trust the AI client as the final control point	Enforce policy in an independent local daemon
Use static, broad permissions	Evaluate each operation in context
Show generic technical prompts	Present business-aware previews
Provide limited decision history	Maintain a local audit trail
Reapprove every routine request	Allow narrow policy rules and expiring approvals
Ignore content sensitivity	Detect likely PII before read content reaches the AI

---

Quick start

Install from the DMG

1. Download the latest PrivacyFence-<version>.dmg from Releases.
2. Drag PrivacyFenceApp.app to /Applications.
3. Install the organization configuration provided by your IT administrator.
4. Authenticate the connectors you want from the PrivacyFence menu bar.
5. Install PrivacyFence.mcpb into Claude Desktop.

Releases are code-signed and notarized by Apple, so this just works — no Gatekeeper warnings, no manual quarantine step. Full installation details are in Technical Reference.

Run from source

git clone https://github.com/andras-tkcs/privacyfence
cd privacyfence
python -m venv .venv
source .venv/bin/activate
pip install -e .

Continue with the organization configuration and connector authentication steps in the Technical Reference.

---

Documentation

• Technical Reference — review model, connectors, policies, installation, configuration, and implementation notes
• Security, Privacy & Compliance — deployment model, data handling, organizational controls, GDPR, and EU AI Act positioning
• Google setup
• Slack setup
• Salesforce setup
• Atlassian setup
• Telegram setup
• Connector QA testing
• Manual pre-release test plan
• Approval window content reference — what each approval dialog shows, grouped by dialog shape
• "Always allow" per-tool reference — what clicking Always allow does, tool by tool
• What Claude knows before an approval prompt
• File type support — attachment previews and PII scanning
• PII detection keywords
• Development vs. installed configuration
• Contributing

---

Status and scope

PrivacyFence has a stable connector and policy interface. It remains under active development, and you should review the limitations and security model before using it with production or regulated data.

Current implementation assumptions:

• macOS host
• local daemon and local approval UI
• MCP-compatible AI client
• per-user connector authentication
• organization-provided OAuth application configuration where required

---

License

Apache License 2.0. See LICENSE and NOTICE.