DOCS
Three surfaces, deployed independently. Start with the workspace connector — it needs read-only OAuth and nothing on anyone’s machine.
step 1 · agentless
Read-only OAuth into Google Workspace or Microsoft 365. This reads the OAuth grants your staff have already authorised, so it tells you which AI tools are in use before anything is installed anywhere.
The connector is read-only. It cannot revoke a grant, and setting a tool to “blocked” creates a policy for future usage rather than removing the app’s existing access.
step 2 · coding agents
A stdio proxy that sits between an AI coding agent and the MCP servers it calls. Each tool call through it is scanned, except an oversized message or batch, which passes through unscanned and is recorded. Every call the gateway can sign is signed locally, by a key that never leaves the machine; where it cannot — standalone mode with no enrolled device, a keypair that will not load, a key registration the backend rejects, or a signing failure on the call itself — the call is still recorded, marked unsigned rather than counted as proven.
python3 --version
# Python 3.10.x or higherThe base package has zero runtime dependencies:
pip install vloex-mcp-proxyFor event sync, heartbeat and enrolment against a backend, install the backend extra:
pip install "vloex-mcp-proxy[backend]"An npm wrapper of the same name is at 0.5.2, which applies the default policy set rather than your organisation’s and has no agent hooks. Use pip.
vloex-mcp-proxy enroll --token your-org-token --backend-url https://api.vloex.comWithout enrolment the proxy still runs and still enforces local policy, but it produces no audit trail — and the audit trail is the part you cannot get anywhere else.
vloex-mcp-proxy discoverDiscovered MCP Clients:
Claude Code (~/.claude.json) (verified)
Servers:
filesystem → npx -y @modelcontextprotocol/server-filesystem /tmp
github → npx -y @modelcontextprotocol/server-github
Cursor (~/.cursor/mcp.json) (verified)
Servers:
postgres → npx -y @modelcontextprotocol/server-postgres
Total: 2 client(s), 3 server(s)Nothing is modified until you drop --dry-run. The rewrite wraps each server’s command so calls route through the proxy.
vloex-mcp-proxy discover --rewrite --dry-run # show the diff
vloex-mcp-proxy discover --rewrite # apply itRestart the client afterwards. Tool calls then appear in the console under Activity, and the wrapped servers show under MCP.
step 3 · browser
Captures each prompt before the request leaves the browser, and the response when it returns. Tested on ChatGPT, Claude, Gemini and Perplexity; set up for 82 more AI providers. Published on both stores:
For fleet rollout there are guides for Chrome Enterprise policy, Microsoft Intune and Jamf; none has been tested on a managed fleet yet.
The extension works without a backend. If it cannot reach one it degrades rather than blocking anybody’s work — and it queues events to send once it can reach one.
step 4 · prove it
The gateway signs every event it can on the device and records the rest as unsigned. The extension in the stores does not sign yet, so its events are recorded as unsigned, and workspace OAuth discovery carries no device signature. The day’s signed events are folded into a Merkle root with a signed checkpoint. Export a sealed bundle from Evidence and verify it offline:
python3 vloex-verify.py bundle.zipThe verifier is a single file with no dependency on us. Fetch the key set independently rather than trusting the copy inside the bundle — verifying against a key that ships in the thing being verified proves nothing.
Try it on a sample bundle without an account.
stuck somewhere
Book a 30-minute call and we will set it up with you, then show you how to export a signed bundle once your events are sealed.