Privacy architecture
Confidentiality built into the architecture.
Not a promise in a policy. There is nowhere on our side for your documents, prompts or keys to land, and you can check the code yourself.
Where your document goes
Opened and edited on your computer. Sent only to the AI you chose.
With the desktop app, the call goes straight from your computer to OpenAI. In the browser it passes through an encrypted tunnel: the connection to OpenAI is secured inside your browser, so our relay only moves bytes it cannot read. Requests are sent with storage turned off.
Requests go directly from your device to OpenAI, Anthropic, Google or the provider you use, under your own key. We are not in the path at all.
Firms run the whole platform on their own infrastructure, under their own domain, sign-in and AI credentials. Your firm is in the path, never us.
What we never hold
We cannot hand over what we never held.
Our servers know who you are and which plan you are on. That is all.
No documents. There is no upload endpoint. Files are opened, indexed and edited on your device.
No prompts or AI responses. No code writes them to disk, database or logs.
No AI keys. Your provider keys and ChatGPT sign-in stay encrypted on your device.
No training. Your work never trains any model, ours or anyone else's.
Your keys, not ours
Your security key unlocks your records. We hold no part of it.
Everything you store is encrypted on your device under a key derived from your own passkey or security key, which never leaves that hardware.
One key, several devices. Enrol a laptop and a security key and both open the same records. Losing one revokes it, not your data.
Firm content stays the firm's. Shared templates use a firm key that your firm's server distributes without being able to read.
No vendor recovery. We cannot reset a key we never had. Firms can keep their own break-glass key; solo users should enrol a second key.
Check the code
Every release is published so anyone can verify it.
The browser runs code we serve, so we make it impossible to change that code quietly.
Signed and logged. Each release is published as a fingerprint of every file, signed and recorded in a public log we cannot edit.
Checkable by anyone. One command compares what a live site serves with what was published.
Checked by your firm. A firm's own server re-checks continuously and refuses to unlock records for a build that does not match.
Apps you connect
When you bring in another AI app, you decide what it sees.
Claude and other AI apps on your computer can work with your documents through the Inchambers desktop app.
Local only. The app talks to Inchambers on your own computer. Nothing passes through our servers.
That app sees what it asks for. Results go to the app you connected and its AI provider, under their terms. This is the one path where data leaves our boundary, and only because you connected it.
You stay in control. Edits arrive as tracked changes under the app's name, every call is logged on your computer, and firms can switch it off.
Limits we state plainly
What this architecture does not protect against.
Your AI provider
The model you choose reads what you send it, under that provider's terms.
A compromised device
Malware on your computer can use your keys while you work. Encryption cannot stop that.
Support tickets
A file you attach to a support ticket reaches us. Do not attach privileged material.
Common questions
Is Inchambers in the path between me and the AI?
With your own AI key or your firm's own servers, no: requests go straight to the provider. With your ChatGPT subscription and the desktop app, the call goes from your computer to OpenAI. In the browser, it passes through an encrypted tunnel we cannot read.
Do you store my documents, prompts or AI responses?
No. There is no endpoint that accepts a document and no code that writes prompts or responses to disk, database or logs.
Can you recover my encrypted records if I lose my security key?
No. We never hold your key, so we cannot reset it. Firms can set up a break-glass key of their own; solo users should enrol a second security key.
Do you train AI on my work?
No. Your work is never used to train or improve any model, ours or anyone else's.
Privacy you can explain to a client.
Tell us your confidentiality and data-residency requirements, and we will map them to the right setup.
