Comparison
AI routers vs attested providers: who can see your prompt
A routing service sits between you and the model. The question is who decrypts your request on the way. This table states what each party publishes or what this directory’s data says; anything else reads “Not stated”.
OpenRouter row summarized from its own documentation
The routing layer, side by side
| Router | Sees plaintext? | Attestation of the router itself | E2EE to the enclave | Retention policy label | Notes / sources |
|---|---|---|---|---|---|
| OpenRouter ↗External routing service | Yes — it screens inference requests, including ZDR traffic, in flight, and samples a small number of prompts for anonymous categorization. | Not stated | Not stated | Per endpoint: ZDR, Safety Retention, or Retention and Training; enforcement covers provider routing for inference only, not plugins/tools. | When it cannot ascertain a provider's data policy it assumes the endpoint retains and trains. Nov 22, 2026: three providers off-boarded and 125 endpoints removed from ZDR eligibility after researchers reported policies not upheld.ZDR docsData collectionNov 22 announcement |
| TinfoilCatalog provider | Not on attested routes, inferred from the directory tier TEE-Terminated TLS: TLS is checked to end inside the TEE. | Not stated | No separate payload key — TLS ends inside the TEE instead. | Not stated | Router enclave at inference.tinfoil.sh; this site's monitor checks it against Tinfoil's signed router release Model enclaves, e.g. the DeepSeek V4.1 Flash replicas; the router attests them, the monitor doesn't check them directly GPUs run in confidential-computing mode; the enclave refuses to start otherwise, per Tinfoilverify panel |
| NanoGPTCatalog provider | Likely yes, inferred from the directory tier TEE-only Gateway: gateway TLS is not shown to end inside the TEE. | Verifier checks reachable NanoGPT TEE model hardware quotes through the authenticated attestation endpoint. NanoGPT is classified as tee-only because gateway TLS and requester nonce binding are not verified. | Not stated | Not stated | TEE models prefixed TEE/; per-request ECDSA signatures verifiable independently of NanoGPT. NanoGPT's docs verify NVIDIA evidence via NRAS but don't name the GPU model; TEE routes are served by upstream providers such as Tinfoil and RedPill.verify panel |
| RedpillCatalog provider | Not the payload, inferred from the directory tier Attested E2EE: the client encrypts to the attested enclave key. | Not stated | Provider-declared capability: ECDH. | Not stated | Attestation anchored to Phala's blockchain for public verifiability RedPill's docs list NVIDIA H100/H200 confidential-computing GPUs across its GPU TEE upstreams (Phala, NEAR AI, Chutes, Tinfoil); the GPU model varies by routeverify panel |
| ChutesCatalog provider | Likely yes, inferred from the directory tier TEE-only Gateway: gateway TLS is not shown to end inside the TEE. | Not stated | Not stated | Not stated | Available on AMD-based Bittensor miner nodes Available on Intel-based Bittensor miner nodesverify panel |
| NEAR AICatalog provider | Not the payload, inferred from the directory tier Attested E2EE: the client encrypts to the attested enclave key. | Not stated | Provider-declared capability: X25519-XCHACHA20-POLY1305. | Not stated | TLS terminates inside the TDX VM — prompts never leave the enclave in plaintext 8× NVIDIA H200 GPUs per node with GPU memory encryption and confidential computing, per NEAR AI's private-inference docsverify panel |
| PPQ.AICatalog provider | Not the payload, inferred from the directory tier Attested E2EE: the client encrypts to the attested enclave key. | The verifier checks the PPQ router and HPKE key for each current private model. Tinfoil backend checks are reported separately, and the PPQ-to-backend route remains unverified unless a request-bound route receipt is exposed. | Provider-declared capability: HPKE. | Not stated | Private inference is accessed through PPQ's private-mode proxy. The verifier checks the PPQ router bundle and attested HPKE key for each private model, then verifies the corresponding Tinfoil model backend enclave.verify panel |
| Venice.aiCatalog provider | Not the payload, inferred from the directory tier Attested E2EE: the client encrypts to the attested enclave key. | Not stated | Provider-declared capability: ECDH. | Not stated | The model attestation response exposes a raw Intel TDX quote and verifier nonce Venice's docs describe NVIDIA Confidential Computing GPUs and return NVIDIA GPU evidence with the attestation; Venice doesn't publish the GPU model. TEE models are operated by NEAR AI Cloud and Phala Network.verify panel |
Catalog rows reuse this directory’s provider data: tier classifications are the monitor’s, E2EE entries are provider declarations. Retention labels are not part of the catalog, so they read “Not stated” rather than being guessed.
Questions to ask any router
- Where does TLS terminate — does the router itself hold the key that decrypts your prompt?
- Is the routing layer itself attested, with a signed quote you can verify?
- Does its retention label cover plugins and tools, or inference routing only?
- Can you choose the upstream provider yourself, or is it chosen for you?
- Can you seed your own canary prompts to detect retention independently?