Run models insidesealed hardware.
Powered by
- Qwen
- DeepSeek
- GLM
- Kimi
- gpt-oss
- Tencent Hy
Same key. One suffix.
Add :incognito to a model id. Your key, SDK and credit work unchanged.
Incognito guideEvery Incognito model, one list
Each runs on Phala hardware and carries the shield badge.
One line changes
Point the model at its :incognito id. Nothing else moves.
No quiet fallback
If Phala cannot serve it, the call fails. Never rerouted.
Sealed from the host.
Intel TDX and NVIDIA confidential GPUs seal memory from the machine's host and operator.
Where your model runs
Deference and its routing provider carry each call to Phala.
Never stored as text
Billing keeps model, tokens and cost, never your words.
How each call travels
Each hop runs over TLS, and Phala seals the memory.
Check it yourself.
Each call names the provider that served it. Phala publishes its hardware attestation.
Served by Phala, on every call
Activity shows the provider reported for each Incognito request.
Modelqwen/qwen3.8-27b
HardwareIntel TDX · NVIDIA confidential GPU
Published byPhala
Attestation, published by Phala
Phala's public report covers the enclave hardware, not one call.
Plain status, no checkmark
Status reads provider reported, exactly what the route tells us.
On record, not on display.
Activity keeps model, tokens and cost for every call. Never the prompt or response.
ActivityYour agent goes incognito too.
Hand your coding agent an :incognito model. Deference MCP runs it the same way.
Deference MCPQuestions
ProductWhat is Deference Incognito?
A route on the Deference API. Add :incognito to a supported model id and the call runs on Phala's confidential-computing hardware, with zero data retention routing and no fallback to another provider.
PrivacyWho can see my prompt?
Phala's hardware protects model memory from its host and operator. Deference and its routing provider process prompt and response text in transit. Deference does not store that text.
ModelsWhich models run in it?
Phala-served models from Qwen, DeepSeek, GLM, Kimi, gpt-oss and Tencent Hy. The models page lists every :incognito id with its price.
BillingDoes it cost more?
Each Incognito model has its own price, set from Phala's rate and paid from the same INFERENCE credit. Every model page shows it before you call.
ChecksHow do I check it?
Activity names the provider that served each call. Phala publishes a hardware attestation for each model; it covers the hardware, not one call.