SayGM
A privacy-preserving AI gateway built on Bittensor subnet SN28, operating within an Intel TDX Trusted Execution Environment. It supports Claude, GPT, Gemini, and open-source models, provides remote attestation, and prevents operators from reading user request content.
Timeline (last 90 days)
Engy (SN53) and SayGM (SN28) are two subnets based on Bittensor.
SayGM is a drop-in gateway that supports Claude, GPT, Gemini, and a range of open-source models, while exposing three API types: OpenAI-compatible, Anthropic-native, and Gemini-native.
The SayGM gateway runs inside an Intel TDX Trusted Execution Environment (TEE), with TLS terminated within the enclave, so that routing and guardrail logic occurs in a place where neither the operator nor the host machine can read the requests.
SayGM provides a remote attestation endpoint accompanied by hardware-signed evidence.
Confidential models in SayGM with the -tee suffix (such as kimi-k3-tee) are open-weight models running inside a TEE, where prompts remain sealed from start to finish.
SayGM's OpenRouter comparison page shows that, after factoring in OpenRouter's 5.5% top-up fee, the average savings across the nine models compared is approximately 29%.
For scenarios using frontier models, providers such as Claude, OpenAI, and Google still see users' prompts in plaintext; SayGM's sealing only covers the gateway layer, not the model providers.
Decision recommendations: choose Engy when you need to know precisely which open-source model responded (e.g., evaluations, long-running agent runs); choose SayGM when using Claude/GPT/Gemini and wanting the routing layer sealed; choose SayGM's -tee models when using open-source models and requiring the strongest privacy.
Neither solution eliminates trust; rather, both shift trust to a place that users can at least partially verify.
All Facts (9)
Engy (SN53) and SayGM (SN28) are two subnets based on Bittensor.
50%UnverifiedSayGM is a drop-in gateway that supports Claude, GPT, Gemini, and a range of open-source models, while exposing three API types: OpenAI-compatible, Anthropic-native, and Gemini-native.
50%UnverifiedThe SayGM gateway runs inside an Intel TDX Trusted Execution Environment (TEE), with TLS terminated within the enclave, so that routing and guardrail logic occurs in a place where neither the operator nor the host machine can read the requests.
50%UnverifiedSayGM provides a remote attestation endpoint accompanied by hardware-signed evidence.
50%UnverifiedConfidential models in SayGM with the -tee suffix (such as kimi-k3-tee) are open-weight models running inside a TEE, where prompts remain sealed from start to finish.
50%UnverifiedSayGM's OpenRouter comparison page shows that, after factoring in OpenRouter's 5.5% top-up fee, the average savings across the nine models compared is approximately 29%.
50%UnverifiedFor scenarios using frontier models, providers such as Claude, OpenAI, and Google still see users' prompts in plaintext; SayGM's sealing only covers the gateway layer, not the model providers.
50%UnverifiedDecision recommendations: choose Engy when you need to know precisely which open-source model responded (e.g., evaluations, long-running agent runs); choose SayGM when using Claude/GPT/Gemini and wanting the routing layer sealed; choose SayGM's -tee models when using open-source models and requiring the strongest privacy.
50%UnverifiedNeither solution eliminates trust; rather, both shift trust to a place that users can at least partially verify.
50%