StackJack 0.1.0.12768+21eadb2
StackJack 0.1.0.12768+21eadb2
This release expands AI Governance with three new ways to oversee AI use: an AI gateway for staff AI traffic, an MCP proxy for remote MCP servers, and Claude hooks. It also adds approval-based changes in vendor AI accounts and the option to use your own storage as the main store. Several fixes address incorrect plan display, false scan-coverage alerts, and Microsoft Graph setup guidance.
New features
- AI gateway. Staff AI traffic to OpenAI, Azure OpenAI and Anthropic can now go through StackJack, so your DLP rules check prompts and answers as they happen.
- Rules can block, warn or mask. Masking is available for OpenAI and Azure OpenAI. Anthropic traffic is inspected but not changed.
- Streamed answers are inspected as they arrive and can be stopped partway through.
- The new AI gateway page under AI Governance settings is where you turn this on, add vendor keys, manage personal keys and set budgets.
- Personal gateway keys. Each staff member gets their own gateway key.
- The key is shown once, with ready-to-copy install commands.
- Staff can see their key card on their own self-view page.
- Generated endpoint settings now include the gateway address.
- Gateway budgets and spend. You can set spending budgets for gateway traffic.
- Administrators and auditors see a budgets band on the spend page.
- Alerts are raised as budgets are reached or calls are refused.
- Proceed after a warning. When a rule warns rather than blocks, the staff member can follow a link to proceed. They are then allowed to continue for 15 minutes.
- Send staff straight to the vendors. A bypass option sends staff traffic directly to the AI vendors instead of through the gateway. Turning it on requires a typed confirmation.
- MCP proxy. Administrators can add approved remote MCP servers on a new MCP proxy page.
- Each server can sign in in one of three ways: no sign-in, one stored header, or each person signs in with their own account. Stored values go into a password box and are never shown again.
- Each server has a tool list policy. Servers can be suspended, resumed and removed.
- Tool call arguments and results are checked against your DLP rules. Matching content in tool results can be masked before it reaches the AI.
- Proxied servers are matched to their approved inventory item.
- An alert is raised when a proxied server keeps failing.
- Personal MCP connections. When a server requires each person to sign in, staff connect their own account through a consent page. The page names whose key the connection is for. Staff can review their connections in a new "Your MCP server connections" card on their self-view.
- Claude hooks. A new Claude hooks page lets you connect Claude hooks to StackJack. Claude tool activity can then be checked against an approved-tool policy and your rules, and allowed or denied.
- Vendor changes. AI Governance can now make changes in Anthropic, OpenAI and GitHub Copilot accounts, once you turn that on for each connection.
- A new Vendor changes tab shows the connection switches, changes waiting for approval and recent changes.
- Each change has a detail page where you can approve it with a typed phrase, reject it, or make the same change again.
- You can start a change from a staff row, an inventory key, the Copilot item and the vendor organizations section.
- The setup pages for these connectors now explain that StackJack's own tools only read. AI Governance makes changes only when you turn that on for the connection.
- Own storage as the main store. You can choose your own storage as the main store for AI Governance content and see its current state. Retention, purge, export and replay follow the main store. An alert is raised if the storage buffer fills up.
- New help pages cover the AI gateway, personal keys and budgets, the vendor bypass, the MCP proxy, Claude hooks, vendor changes and own storage as the main store.
Improvements
- The AI Governance home band now shows separate counts of blocked and warned prompts for the last 7 days. Each count links to the matching findings.
- The Findings page has a new "What StackJack did" filter, so you can list findings by the action taken.
- When the Anthropic activity feed is refused for a missing permission, the message now explains the fix. Turn on the Compliance API for the organization, then create a new Admin API key.
- Staff or domains you have excluded from content capture are recorded with facts only on the gateway, MCP proxy and hook lanes. Their prompt text is not read or kept.
Bug fixes
- Organizations using the admin connectors included with AI Governance could see a Free plan on the Connectors and Billing pages. Tool lists could also hide tools the connector was entitled to. These pages now show "Included with AI Governance", and tool lists match what the connector can run.
- Fixed recurring "Scan coverage lost" alerts on StackJack MCP tool calls. These alerts were raised even though the scan had not actually failed.
- AI-assisted content checks on long or dense text could fail and fall back to rules only. That content is now split into smaller parts and checked in full.
- Leaving a Portal page while its guided tour was starting could end the Portal session. This no longer happens.
UI & UX changes
- DLP rule pages now show which actions each lane supports.
- The Microsoft Graph card and setup guide no longer say GDAP customers come along automatically.
- They now explain that each customer tenant needs a one-time admin consent, using "Authorize a customer tenant" on the Microsoft Graph card.
- "Authorize all customers" is available when your organization connects its own app registration.
Fixed customer-reported issues
- Microsoft Graph connections failed with a consent error on GDAP customer tenants (
01M63GB3PHC6BQ5N0Y8WAVFMWY). The Microsoft Graph card and setup guide now explain that each customer tenant needs a one-time admin consent and point to the button that grants it.