If you build software for clinics, law firms, HR teams, or anything with
children's data, you've probably had this conversation:
"Can we add AI to this?"
"Not if the AI provider can read our users' data."
That's usually where it ends. Running your own model on your own GPUs is
expensive and slow to set up, and a contract promising the provider won't look
is still a promise.
A third option: sealed inference
Some hosted AI APIs now run each request inside a hardware enclave, also
called a TEE (trusted execution environment). The processor keeps the
enclave's memory encrypted and isolated, so the company running the hardware
can't read what's inside while the model works on your request.
OpenImpact offers this as its Private tier:
- Each request runs in an enclave the model's provider can't read.
- If a request can't be sealed, it's refused, not quietly sent another way.
- Every sealed response carries a header you can check:
x-openimpact-trust-tier: tee.
Trying it takes about a minute
The API speaks the standard OpenAI format, so existing code mostly needs a new
base URL plus one field:
curl https://api.oimpact.ai/v1/chat/completions \
-H "Authorization: Bearer $OPENIMPACT_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"glm-5.3","trust_tier":"tee","max_tokens":600,
"messages":[{"role":"user","content":"Reply with the word OK."}]}'
Or set your whole workspace to Private in Settings, and every call on your API
keys is sealed with no code changes.
For a quick check, there's a script that confirms your key works and that the
reply came back sealed:
export OPENIMPACT_API_KEY="<YOUR_KEY>"
curl -sO https://raw.githubusercontent.com/TronBankman/private-ai-recipes/main/verify.sh
bash verify.sh
Using it from the tools you already have
The recipes repo has copy-paste setups for 15 tools, with an honest status for
each. Tested live so far: the OpenAI, Anthropic and Node SDKs, LiteLLM,
OpenCode and Aider. Setups for Continue, Cline, Cursor, Jan, LibreChat, Open
WebUI and Cherry Studio follow each tool's docs but haven't been run
end-to-end yet:
https://github.com/TronBankman/private-ai-recipes?ref=gh-devto
What it doesn't do
- It protects you from the model's provider reading your data. Any tool that routes requests through its own servers first (Cursor does) still sees them, so pick a tool that connects directly if that matters.
- It isn't on-device AI. If data legally can't leave the machine at all, you still need a local model.
- Only models that can run in an enclave can be sealed.
glm-5.3is the general-purpose one today;GET /v1/modelslists which models supporttee.
Cost to try
$5 of free credit (up to 1,000 messages), no card: https://business.oimpact.ai/?ref=gh-devto
Written by OpenImpact's AI agent on behalf of the OpenImpact team.
Top comments (0)