Confidential LLM inference · OpenAI-compatible

The LLM API that can’t read your prompts. Check it yourself.

One API for open-weight models. Each model runs in a hardware-protected confidential environment, so no one — including us — can read your prompts. You can verify this before sending any data.

Have an invite?

No invite yet? Join the waitlist · Watch the explainer

OpenAI-compatible · streamingAMD SEV-SNP · Intel TDX · NVIDIA CCOpen-source Gatekeeper
python · your existing openai SDK1 line changes
from openai import OpenAI
client = OpenAI(
base_url="https://api.router.superprotocol.com/v1", # the only line that changes
api_key=os.environ["SUPERPROTOCOL_API_KEY"],
)
client.chat.completions.create(
model="meta/llama-3.2-3b-instruct:tee",
messages=[{"role": "user", "content": "Summarise this patient record…"}],
)

See how it works.

Super Swarm, the confidential infrastructure underneath, and how Confidential Router uses it to serve models no one can read.

Hardware instead of promises.

Models run in a confidential execution environment (TEE) on CPU and GPU. The hardware enforces the protection.

Isolated execution

The CPU (AMD SEV-SNP or Intel TDX) and the GPU (NVIDIA Confidential Computing) create an isolated environment. Nothing outside it can access what runs inside.

Encrypted memory

Memory is encrypted with keys that never leave the chip. Data moving between CPU and GPU is encrypted too.

Who can’t read your data

  • The cloud operator
  • The host operating system
  • Other tenants
  • Super Protocol

Proof flows one way.

The confidential environment publishes signed proof of what it runs. Gatekeeper checks it on your machine and decides whether to send data.

/ BEFORE · AN ORDINARY API
Your appprompt · plaintext
Provider APIprompt · readable by the provider

The provider receives the prompt in plaintext. Only a privacy policy protects it — for regulated data, that often fails security review.

/ WITH CONFIDENTIAL ROUTER
Confidential Router: request path and evidence pathLeft, your machine: your app sends its request to Gatekeeper, a local proxy. Centre, the public internet, which carries only TLS ciphertext. Right, a hardware-protected confidential environment spanning CPU and GPU (AMD SEV-SNP or Intel TDX with NVIDIA Confidential Computing), containing the API endpoint, the model server and the model weights in one protected memory space; prompts are readable only on your machine and inside that environment. Step 1: Gatekeeper fetches the signed proof the environment publishes. Step 2: it confirms genuine TEE hardware running approved code. Step 3: it checks the signature and freshness. Step 4: it confirms the live connection is the one the evidence describes, and only then forwards the request to the endpoint. The verdict ends at Gatekeeper; nothing is sent back to the endpoint. Outside it, the cloud operator, the host OS, other tenants and Super Protocol cannot read its memory.OUTSIDE · CANNOT READ YOUR DATACloud operatorHost OSOther tenantsSuper ProtocolYOUR MACHINEprompt readable hereYour app or agentprompt · plaintextlocalhostGatekeeperlocal proxy · runs on your side2hardware3signatureverdict: ADMITverdict ends here —never sent to the endpointPUBLIC INTERNETTLS ciphertext onlylive connection must match the evidence4request → endpoint · ← response · TLSCONFIDENTIAL ENVIRONMENT (TEE)CPU: AMD SEV-SNP / Intel TDX · GPU: NVIDIA CCencrypted memory · encrypted CPU↔GPU transfersAPI endpointOpenAI APIModel serverconfidential GPUModel weightsopen-weightone protected memory space · prompt readable only hereSigned proofsigned proof · publishedwhich codeon which hardware1fetch the published evidenceConfidential Router: request path and evidence pathTop, your machine: your app sends its request to Gatekeeper, a local proxy. Middle, the public internet, carrying only TLS ciphertext. Bottom, a hardware-protected confidential environment spanning CPU and GPU (AMD SEV-SNP or Intel TDX with NVIDIA Confidential Computing), containing the API endpoint, the model server and the model weights in one protected memory space, plus the signed deployment evidence it publishes. Step 1: Gatekeeper fetches the evidence. Step 2: it confirms genuine TEE hardware. Step 3: it checks the signature and freshness. Step 4: it confirms the live connection matches the evidence before forwarding. The verdict stays at Gatekeeper and is never sent to the endpoint. Outside it, the cloud operator, the host OS, other tenants and Super Protocol cannot read its memory.YOUR MACHINEprompt readable hereYour app or agentprompt · plaintextlocalhostGatekeeperlocal proxy · runs on your side2hardware3signatureverdict · stays herePUBLIC INTERNETTLS ciphertext onlylive connection mustmatch the evidence4CONFIDENTIAL ENVIRONMENTCPU: SEV-SNP / TDX · GPU: NVIDIA CCAPI endpointOpenAI APIModel serverconfidential GPUModel weightsopen-weightone protected memory spaceSigned proofsigned proof · published1OUTSIDE · CANNOT READ YOUR DATACloud operatorHost OSOther tenantsSuper Protocol
Trust boundary
Request and response, over TLS
Proof and verdict
Cannot read your data
  1. 1Fetch the proof
  2. 2Check the hardware
  3. 3Check signature and freshness
  4. 4Match the connection, then send
THE RULE

We publish proof. Your software decides. We never learn whether or when you checked us.

Check first. Then send.

Gatekeeper is an open-source local proxy for macOS, Linux and Windows. It sends nothing until four checks pass.

  1. 1

    Fetch the proof

    Download the signed proof: which code runs, on which hardware.

  2. 2

    Check the hardware

    Confirm it is backed by a TEE quote signed by the chip vendor.

  3. 3

    Check signature and freshness

    Confirm the proof is signed and current.

  4. 4

    Match the connection

    Confirm you are connected to that exact environment. Only then send data.

DEFAULT · FAIL-CLOSEDA check fails — nothing is sent.

Your app gets an error.

OPTIONAL · FAIL-OPENTraffic continues without proof.

Useful for testing.

Root of trust: the chip vendor, not Super Protocol. You can pin a specific deployment and add your own rules. Any tool can check the proof, not only ours. There is no certificate to paste: Gatekeeper accepts a Swarm cloud only when the chip vendor’s report and a measurement Super Protocol signed in its public registry both vouch for it. For a local or test build that registry has not signed, the optional block adds the cloud’s certificate authority, served from this page — it should print sha256:2d9c20a6…54c276; if it prints anything else, stop.

installmacOS · Linux · Windows
curl -fsSL https://github.com/Super-Protocol/confidential-router/releases/latest/download/install.sh | sh
set up · four commandsfresh machine, nothing configured
gatekeeper init
gatekeeper endpoint add router --upstream https://api.router.superprotocol.com --listen 127.0.0.1:8787
gatekeeper endpoint trust add router --from-upstream # review the report, then confirm
gatekeeper run # --headless if you have no terminal
Optional: test or unsigned builds

Needed only when the cloud’s measurement is not signed in the Super Protocol registry — a local or test build — and Gatekeeper denies with “not in the Super Protocol trusted registry”. A running Gatekeeper asks the registry again within about fifteen minutes, so a cloud signed later is admitted without a restart.

test or unsigned buildeither line
gatekeeper trust measurements add --from-upstream router # accept the measurement, or:
curl -fsSL https://router.superprotocol.com/swarm-root.pem | gatekeeper trust roots add swarm-prod --pem-file -
gatekeeper · verdictlocal
$ gatekeeper verify router
ADMITTED — verified and allowed by every policy
Endpoint router
Evidence kind DeploymentEvidence
Issued 3s ago
Trusted root swarm-prod (sha256:2d9c20a6…54c276)
Observed TLS leaf sha256:2c88a1a6…0dad00
Signed certFingerprint sha256:2c88a1a6…0dad00
evidenceDigest sha256:b76515af…027d24
Pinned for this endpoint yes
Policies
(built-in) gatekeeper.default allow
Computed on your machine. We are not told the result.

Attestation proves

  • Which code is running.
  • That it runs on genuine TEE hardware.
  • That you are connected to it.

It does not prove

  • Model quality.
  • What happens to data on your side.
  • Anything, if you skip Gatekeeper: your data is still protected, but you have no proof.

Change one line.

Change the base URL to https://api.router.superprotocol.com/v1. Your OpenAI SDK, keys and streaming work as before.

Code examples
curl https://api.router.superprotocol.com/v1/chat/completions \
-H "Authorization: Bearer $SUPERPROTOCOL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/llama-3.2-3b-instruct:tee",
"messages": [{"role": "user", "content": "Summarise this patient record…"}]
}'
Any OpenAI-compatible client works.
from openai import OpenAI
client = OpenAI(
base_url="https://api.router.superprotocol.com/v1", # the only line that changes
api_key=os.environ["SUPERPROTOCOL_API_KEY"],
)
client.chat.completions.create(
model="meta/llama-3.2-3b-instruct:tee",
messages=[{"role": "user", "content": "Summarise this patient record…"}],
)
The official openai package.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.router.superprotocol.com/v1", // the only line that changes
apiKey: process.env.SUPERPROTOCOL_API_KEY,
});
const completion = await client.chat.completions.create({
model: "meta/llama-3.2-3b-instruct:tee",
messages: [{ role: "user", content: "Summarise this patient record…" }],
});
The official openai package for Node.
from openai import OpenAI
client = OpenAI(
base_url="http://127.0.0.1:8787/v1", # your local Gatekeeper
api_key=os.environ["SUPERPROTOCOL_API_KEY"],
)
client.chat.completions.create(
model="meta/llama-3.2-3b-instruct:tee",
messages=[{"role": "user", "content": "Summarise this patient record…"}],
)
Same code, pointed at your local Gatekeeper. Data is sent only after verification.
USD per 1M tokens. Pay as you go, no subscription.
ModelAPI idContextInputOutput
Gemma 2 2Bgoogle/gemma-2-2b-it:tee8K$0.02$0.03
Llama 3.2 3B Instructmeta/llama-3.2-3b-instruct:tee128K$0.03$0.05

:teeThe model runs in a confidential environment.

CatalogueOpen-weight models, from 2B to 3B. DeepSeek and other larger models will follow as capacity grows.

$100 creditLlama 3.2 3B Instruct at an even input/output mix averages $0.04 per 1M tokens — so $100 is about 2.5 billion tokens.

We meter. We don’t read.

What we meter

  • Tokensin / out, per request
  • Costper request
  • ModelAPI id
  • Latency and timingper request
  • Which keysk-tee-v1-…

What we never hold

  • Prompt textnever stored
  • Completionsnever stored
  • Attachmentsnever stored
  • Embeddings contentnever stored

For data that has to pass review.

Agents and assistants that need the real data, not a redacted copy. Open a case to see examples.

Use cases · 01

Software engineering

No need to strip secrets before the agent sees them. Nothing readable leaves the confidential environment.

  • Coding agents over the whole repository, including internal libraries and private dependencies.
  • Service and ops agents that read configs, connection strings, API keys and logs.
  • Debugging on real production data and error traces, without scrubbing them first.

Use cases · 02

Healthcare

Patient data is processed where no cloud, no vendor and no one at Super Protocol can read it — a fact you can prove in a HIPAA or GDPR review, not a clause in a BAA.

  • A physician consults an agent on a specific case: history, labs, imaging reports, current medications.
  • Agents that draft discharge summaries, referral letters and prior-authorization requests from the chart.
  • Medication-interaction and guideline checks against the patient’s actual data.

Use cases · 03

Financial services

Material non-public information never sits in readable memory at a third party.

  • Research and deal agents over pipelines, portfolios and unaudited financials.
  • AML and KYC case review across customer transactions and documents.
  • Credit memos and risk assessments drafted from internal statements.

Use cases · 04

Enterprise analytics

Commercial data stays inside your perimeter technically, not contractually — which is what security review asks for.

  • Analytical agents over revenue, pricing, margins and supplier contracts.
  • Agents connected to internal databases, BI exports and CRM.
  • Strategy and board briefs drafted from plans, forecasts and M&A material.

Questions a sceptic should ask.

Q.01How is this different from a privacy policy?

A policy is a promise about data the provider can read. Here the provider can’t read it — and you can check that yourself.

Q.02What if Super Protocol is compromised?

We can’t read your data either. If the code changes, the proof changes, and Gatekeeper blocks it. We can still stop the service: you get an error, not a silent leak. See the threat model.

Q.03Is it slower or more expensive?

Encryption adds some overhead. Test it on your workload with the starting credit. Prices are in the table above.

Q.04How do I get access?

Access is by invitation, and each invite comes with a $100 credit. No invite yet? Join the waitlist. After the credit, you pay as you go.

Q.05Is it open source?

Yes. Gatekeeper and the docs are on GitHub.

Q.06What if verification fails mid-session?

By default Gatekeeper stops sending and your app gets an error. In fail-open mode, traffic continues without proof.

Your prompts are not our business. See for yourself.

Get an invite. Change the base URL. Run Gatekeeper if you want proof.