DEV Community

Morse
Morse

Posted on

Private, OpenAI-compatible LLM APIs compared: zero retention, TEE and verifiable enclaves

Every fact about each API comes from its own docs, linked inline, checked on November 20, 2026.

If you send prompts to a hosted model, "private" can mean any of three different things:

  • Zero data retention (ZDR): the provider promises, usually in a contract, not to store your prompts or replies. You're trusting a policy.
  • TEE (trusted execution environment): the model runs inside hardware-isolated memory (Intel TDX, AMD SEV-SNP, NVIDIA confidential-computing GPUs) that the operator can't read. You're trusting the chip maker's hardware.
  • E2EE (end-to-end encryption): your client encrypts the request to a key that only the enclave holds, so nothing between you and the enclave can read it. The useful question is which enclave: a gateway's, or the one running the model.

Attestation ties it together. It's a report signed by the hardware vendor that says what code an enclave is running, so you can check it against published source instead of taking the provider's word for it.

The comparison

Morse first (it's ours), then the others by type: routers, enclave hosts, gateways. This isn't a ranking.

API OpenAI-compatible ZDR / no retention TEE E2EE Attestation you can check Models
Morse (ours) Yes Yes, private tier by contract; Morse's enclave keeps no prompt or reply Yes, tee tier Yes, e2ee tier, to the model's enclave Yes, npx @askmorse/verify 87 open-weight models
OpenRouter Yes Yes, per request (zdr: true) Not described on its pages Not described Not described Open and proprietary
TrustedRouter Yes Yes (prompt storage off) Gateway in a TEE; confidential routes via other hosts TLS ends inside the gateway's enclave Yes, image digest at /attestation Open and proprietary
Phala / RedPill Yes Gateway stores no bodies; zdr: true upstream Yes (Intel TDX, NVIDIA H100/H200/B300) To the gateway Yes, @phala/aci-verifier, dstack verifier Open and proprietary (proprietary outside the TEE)
NEAR AI Cloud Yes "No logs of prompts or responses" Yes (Intel TDX + NVIDIA H200) To the model's enclave Yes, open-source verifier + web page Open in TEE; GPT/Claude/Gemini outside it
Tinfoil Yes Yes Yes (AMD SEV-SNP, Intel TDX, NVIDIA H100/H200/B200) To the model's enclave (TLS in enclave + EHBP) Yes, SDKs + tinfoil attestation verify Open-weight only
Prem Yes (gateway or local proxy) Yes, in ZDR mode Yes, in confidential mode (SEV-SNP/TDX + NVIDIA) Yes, in confidential mode (ML-KEM768 + X25519) Yes, Reticle verifier Open-weight only
Privatemode Yes (local proxy) Prompts not stored Yes (SEV-SNP, TDX, NVIDIA H100/B200) To the worker (HPKE) Yes, via the proxy; reference values public Open-weight only

"Not described" means we couldn't find it on that company's own pages. It doesn't mean the feature doesn't exist.

Provider notes

Morse (ours). One OpenAI-compatible endpoint for 80+ open-weight models, "called from Morse's attested enclave, which keeps no prompt or reply" (docs). TLS terminates inside the enclave, and providers see Morse as the customer, so "your account, your card, and your IP address stop at Morse" (inference). You set a privacy floor per request: private (providers bound by contract to keep nothing), tee (the model runs in the provider's enclave) or e2ee (encrypted to a key only the model's enclave holds). Morse "refuses a request it can't serve at that floor rather than sending it lower." The enclave code is open source, and npx @askmorse/verify inference.trymorse.ai checks the running release against it. Pricing is pay as you go, with no subscription.

OpenRouter. One key for open and proprietary models. Its ZDR page defines ZDR as "no routine storage of prompts or responses after processing, apart from the CSAM exception," and you can require it per request with "provider": {"zdr": true}. Provider routing also has data_collection: "deny" ("use only providers which do not collect user data"). The ZDR page notes that self-serve accounts allow "Safety Retention" endpoints by default, so check your account settings.

TrustedRouter. An OpenAI-compatible router at api.trustedrouter.com/v1 with a confidential-only host as well (docs). Its trust page says "Public TLS terminates inside the measured GCP Confidential Space workload," and you can compare the image_digest from /attestation with the published one. Per request you can set "provider": {"min_privacy": "zdr"} or "confidential". The router code is open source.

Phala / RedPill. RedPill's gateway runs in Intel TDX, and "the gateway does not store request or response bodies" (trust boundary). Its E2EE "is decrypted inside the attested gateway" (E2EE docs). GPU-TEE models run on H100, H200 and B300 (Phala). Proprietary models run on providers that aren't attested, so check is_tee per model.

NEAR AI Cloud. It runs on "Intel TDX + NVIDIA GPU TEE" (near.ai). Its E2EE goes all the way to the model: "Only model environments holding the attested model key can decrypt your messages" (E2EE guide). OpenAI, Anthropic and Gemini models are offered in an "Incognito" mode that runs "outside a TEE" (models). The verifier is open source.

Tinfoil. Open-weight models only, at inference.tinfoil.sh/v1. "We do not retain prompt or response content after the response is returned" (privacy). TLS terminates inside the confidential VM, and its EHBP protocol "encrypts HTTP message bodies end-to-end." Its SDKs verify attestation for you. Tinfoil notes that plain OpenAI-SDK calls skip that check and are "not recommended for production."

Prem. Two modes. ZDR mode (gateway.prem.io/openai) is contract-based: "Nobody keeps your prompts, completions, or audio content at any stage of the request" (ZDR). Confidential mode runs through a local proxy: "The SDK or local proxy encrypts the request payload before network egress" (security model), and the Reticle verifier checks the CPU and every GPU.

Privatemode. Built by Edgeless Systems. It's "compatible with the OpenAI and Anthropic APIs" through a local proxy (docs). "Prompts are encrypted by the client-side Privatemode proxy and decrypted by a server-side" proxy on the worker (encryption). It also says "the reference values in the manifest can be derived by anyone" (attestation).

Switching is a base-URL change

All eight speak the Chat Completions format, so the OpenAI SDK you already use keeps working. Here's the Morse version, with a privacy floor on one request:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://inference.trymorse.ai/v1",   # was https://api.openai.com/v1
    api_key=os.environ["MORSE_API_KEY"],           # was OPENAI_API_KEY
)

reply = client.chat.completions.create(
    model="z-ai/glm-5.3",
    messages=[{"role": "user", "content": "Summarise this contract clause."}],
    extra_body={"provider": {"min_privacy": "tee"}},  # only attested hardware may serve it
)
print(reply.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

The same pattern works for the others: swap base_url, the key and the model id, and move any routing options into extra_body (for OpenRouter, {"provider": {"zdr": True}}).

How to choose

  • You need GPT or Claude: OpenRouter, TrustedRouter, Phala/RedPill and NEAR AI route to proprietary models. There, privacy comes from ZDR terms, not an enclave.
  • You want the model itself in attested hardware: Morse's tee and e2ee tiers, Tinfoil, NEAR AI, Prem (confidential mode), Privatemode and Phala's GPU-TEE models.
  • You want encryption that ends at the model, not at a gateway: Morse's e2ee tier, NEAR AI, Tinfoil, Prem and Privatemode.

Whatever you pick, run the provider's verifier in CI rather than once by hand. That turns a privacy claim into a check that can fail.

FAQ

Is zero data retention the same as a TEE?
No. ZDR is a promise not to store your data, usually backed by a contract. A TEE is hardware that stops the operator from reading memory while your request runs, and attestation lets you check which code is running.

Can I keep using the OpenAI SDK?
Yes, with all eight. You change the base URL, the key and the model id. Prem and Privatemode can also run a local proxy that you point the SDK at.

What does "attestation you can check" mean in practice?
A command or library fetches the enclave's hardware-signed report and compares its code measurement with a published build. Examples: npx @askmorse/verify, tinfoil attestation verify, @phala/aci-verifier, NEAR AI's verifier and Prem's Reticle.

Does E2EE mean the provider can't see my prompt?
It depends on where the encryption ends. On RedPill it ends at the attested gateway. On Morse's e2ee tier, NEAR AI, Tinfoil, Prem's confidential mode and Privatemode, it ends inside the enclave that runs the model.

Which of these offer GPT and Claude?
OpenRouter, TrustedRouter, Phala/RedPill and NEAR AI. Morse's API, Tinfoil, Prem and Privatemode serve open-weight models such as DeepSeek, GLM, Kimi, Qwen and gpt-oss.

How do I make sure a request is never served at a lower privacy level?
Set the floor on the request. Morse and TrustedRouter have provider.min_privacy, OpenRouter has provider.zdr, and RedPill has provider.zdr, which fails closed with a 404.

How does Morse's API price requests?
Pay as you go, with no subscription. Each model's price per million tokens is listed on the models page.

Top comments (0)