The firm’s AI agents, on the record.
AI agents put themselves on the register, down to the one running quietly on a single laptop. Each ties back to a named person, and each has an off switch.
Agents
Every individual agent on the register, recorded from its own traffic the moment it first appears. Each row is one agent with one owner, one home, and its own controls.
How Agent Governance works with Proxara.
An employee hands an agent a task and walks away. The agent drifts out of its lane. Here is what happens next.
Agentsfenwick-ridgeEverything in the film is the product: the register, what each agent can reach, the hold, and Rox.
Every agent has an owner and an off switch.
Open any row and the register goes one level deeper: who is accountable, what the agent is allowed to reach, and the controls to slow it, hold it, or stop it outright.
One Claude Code agent, shown exactly as it appears in the console.
Claude Code · ws-9f31c2ab
Observed and recorded on every covered device.
Everything recorded for this agent traces back to envelopes the device signed before they left the machine. The extract bundles those envelopes with the device keys, Merkle proofs, signed anchors, and a verifier that runs offline. An examiner can check it without access to Proxara.
What Proxara covers, and where it stops.
Every mark below is something the product does today, not a plan. Where coverage has a limit, the row says so.
| Surface | On the register | Redacted in transit | Block & limit | Researched |
|---|---|---|---|---|
Tools connected inside managed AI apps Claude Desktop Cursor Gemini CLIWindsurfVS CodeEvery tool in these apps is found on install and lands on the register. The ones running on the laptop itself are blocked or warned by policy. | ||||
| Cloud connectorsProxara Connect, and Claude’s own directoryThe firm’s own systems connect through Proxara Connect, governed end to end wherever the person signs in. Other connectors in Claude’s directory are seen by name, and the firm can turn them off in Claude’s own admin. | ||||
| Coding agents on the command lineClaude CodeCodexTheir settings are read and listed, never rewritten. The agents themselves sit on the register, where they can be limited, blocked, or stopped. | ||||
| Beyond the deviceother makers’ cloud add-ons, and machines without ProxaraCloud add-ons from other AI makers, and machines Proxara was never installed on. Named plainly rather than implied. |
New servers get researched, not just listed.
The moment a server the firm has not approved appears on a device, Rox researches it: real sources, written reasoning, and a call to allow, watch, or block.
- A risk rating of low, medium, or high, with the reasoning written out.
- Links to real sources, so the call can be checked.
- One click to block, and the block reaches every covered device.
- If the firm chooses, a bad verdict holds the server automatically.
No verified maker, and it asks for more of the computer than its job needs. Two public reports of copycat versions under the same name.
Sources
Package registry: publish historyPublic advisory: lookalike namesAssessed by Proxara intelligence.
One page a board can read.
Which agents are running, who owns each one, what they touched, and that all of it was governed. Signed, so it holds up outside the room it was written in.
The person responsible walks in with an account of the firm’s agents, not a promise about them.
Book a demoA one-page summary a board receives.
Northbridge Capital
AI agents on the register9
With a named owner8
Blocked this period1
Ungoverned0
Signed, so it can be checked by an outside examiner.
Where the register leads.
An agent register is only useful next to the things that explain who owns an agent, what it can reach, and where it runs.
Put every agent on the record.
Tell us what the firm runs, and a person will reply in a few hours.
Talk to us