AgentKeyRoute Guard (browser extension)
Much AI use never touches an API: people paste into chat websites. Guard is the AgentKeyRoute browser extension for ChatGPT, Claude, Gemini, Copilot and Perplexity. The same extension can run in four modes, which your admin flips per site and per team from the console.
The four modes
| Mode | What Guard does | Governance |
|---|---|---|
| off | Nothing. | — |
| log | Observes prompts and replies on vendor sites and sends them to POST /v1/log. Nothing is blocked. | Passive audit. |
| inspect | Before each message is sent, Guard calls POST /v1/inspect. Blocked data never leaves the browser; redactable details are swapped for placeholders. The user still talks to the vendor on their own account. | Pre-send policy. |
| proxy | Hides the vendor chat UI and mounts the AgentKeyRoute chat on the same site, routing every request through the gateway on a company key. | Full request and response control. |
Install
- Create an extension key
In the console, open Connect an app → Browser extension (opens in a new tab), choose the profile that decides what may be shared (Internal assistant by default) and click Create key. A team can share one extension key; each person's email identifies them.
- Download the extension
On the same page, download
agentkeyroute-guard.zip. It works in Chrome, Edge, Brave and other Chromium browsers. - Install it
- Unzip the file.
- Open
chrome://extensions(oredge://extensions) and turn on Developer mode. - Click Load unpacked and choose the unzipped folder.
For the whole company, IT can force-install it with browser management policy so it can't be switched off.
- Connect it
Click the AgentKeyRoute icon, enter the gateway address
https://gateway.agentkeyroute.com, the key and your email, then Save and test. - Pick the mode per site
In the console, under Browser extension (opens in a new tab), set the mode for each site your team uses: ChatGPT, Claude, Gemini, Copilot, Perplexity. Start with log to build an inventory; move to inspect or proxy as policy hardens.
What people see in inspect mode
| Outcome | What happens |
|---|---|
| Nothing sensitive | The message is sent as normal. |
| Details replaced | Names, emails or patient details are swapped for placeholders in the box; the person reviews and presses send again. |
| Not sent | Blocked data, such as card numbers or passwords, never leaves the browser. |
Every check is recorded in Requests as "Guard · <mode> · <site>". In inspect mode no model is called by AgentKeyRoute; the user still talks to the vendor on their own account. Under the hood it uses POST /v1/inspect, which you can call from your own tools too:
curl https://gateway.agentkeyroute.com/v1/inspect \
-H "Authorization: Bearer <your API key>" \
-H "Content-Type: application/json" \
-d '{"destination":"chatgpt.com","user":"[email protected]","text":"Summarise this note for John Smith"}'Proxy mode in detail
When a site is set to proxy, Guard hides the vendor chat UI and mounts the AgentKeyRoute chat on the same page. Every message routes through the gateway on a company key, using the models and provider connections the profile allows. Cost and usage are attributed to the person (by email) and the workload. Users do not need a vendor account; the vendor login is bypassed.
Proxy mode is strongest when: you want one audit log for every AI session; you want users to stop using personal accounts entirely; or the regulatory story needs provider choice and region control on top of inspection.
Log mode in detail
Log mode is the lowest-friction way to find out how your team is using public AI. Guard observes prompts and replies on vendor sites and sends them to POST /v1/log. The profile decides whether content is captured in full, redacted or kept as metadata only. No policy blocks anything, so nobody's workflow changes.
Most organisations start with log for two weeks, review what shows up in Requests, and then move each site to inspect or proxy with confidence.