Before connection
Choose the execution path, not the model.
These are implemented integration paths, not a claim that every agent or OS version has passed a live acceptance test. All first-release hooks and SDKs are cooperative: tools outside them are not protected.
On narrow screens, scroll the table horizontally to read approval and execution-receipt support. Keyboard users can focus the table and use the arrow keys.
| Runtime / path | OS & requirements | Permissions | Approvals | Execution receipts & limits |
|---|---|---|---|---|
| Claude Code native connector | Windows, macOS, desktop Linux; x64 / ARM64 builds | Workspace connection permission; writable Claude hook settings; DPAPI / Keychain / Secret Service. Linux needs secret-tool and an unlocked keyring. | Claude native prompt or website review, subject to workspace reviewer policy | Supported post-tool hooks report outcomes. Verify registration and receipt arrival in your host version; permission is not execution. Restart Claude after setup/removal. |
| MCP enforcement wrapper | Local stdio MCP server on a supported connector host | Control of client/server launch configuration; scoped coding credential | Website review; resume the exact action after approval | Coverage only for wrapped calls. A tool return or host receipt does not independently prove side effects. Not a general remote-MCP proxy. |
| OpenAI, Ollama, custom application SDK | Python / JavaScript runtime; application-owned tool dispatcher | Code changes at every protected tool execution point; scoped coding credential | Website review through the adapter; no automatic native host prompt | Depends on the adapter's result reporting. Direct provider calls and unwrapped tools are outside coverage. Model provider is independent of enforcement. |
| Other coding-agent hosts (including Cursor / Copilot) | Host/version-specific extension point required | Integration code or supported hook registration | Do not assume native approval support | No universal turnkey compatibility guarantee. Validate the actual host lifecycle before rollout; a supported tool-call format alone is not a complete integration. |
| Forensics MCP / Investigator | MCP client or browser; existing Pramagent records | Read-only forensic access within the workspace | Cannot approve or execute tools | Reads available evidence only. Does not collect external logs, create execution receipts or enable enforcement. |
First successful check
- Review data handling and pilot expectations, then create a new code in guided setup.
- Register the integration and ask it to read a synthetic README.md. Match the fresh action to the exact credential created in this setup.
- Use the dashboard's policy-only allow/block checks: they execute no tools. Open the resulting decision record. They test policy reachability, not local hook interception.
- Open the session investigation and inspect execution coverage. A host receipt is self-reported; a missing receipt means the outcome is unknown.
Unsigned preview downloads
Windows publisher signing and macOS signing/notarization are pending. Do not bypass browser or operating-system warnings. SHA-256 verifies bytes, not publisher identity. If your organization requires signed software, use the no-install sample while signing is pending; do not treat this page as approval to install.
Control my agent · Investigate existing activity · Ask about your runtime