This server exposes MCP tools for encrypted/private computation: status checks, op discovery, sealing values, running FHE operations, and revealing final results.
Check private-compute status and list available private ops (add, sub, mul, div, mean, concat, compare, etc.).
Seal plaintext values (int, float, string, binary) into ciphertext handles (ct_...), optionally returning raw ciphertext or using public-key encryption.
Run private operations on sealed handles and/or plaintext inputs, with optional final-result decryption.
Decrypt sealed handles to reveal plaintext only when explicitly requested.
Use one-shot fhe_private_eval to seal inputs, compute an operation, and optionally reveal only the final answer.
Work with encrypted integers, floats, strings, and binary data; no model inference, SQL, or retrieval.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@aura Privately add 25 and 17."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
FHE is the foundation. Aura AI is the application. MCP connects your assistant to encrypted tools.
FHE lets software calculate with data in its encrypted form. Aura's flagship AI application gives that foundation a practical purpose: useful AI with a defined privacy and key-custody model. MCP is the connection from your assistant to the capabilities exposed here.
Aura reports GPT-OSS-20B encrypted inference at 20+ tokens/second on one RTX PRO 6000 Blackwell GPU, with prompt read-in below 2.3 seconds. This is the website's internal benchmark, not a speed measured by this MCP or an independent validation. Benchmark source and scope.
Three ways to experience Aura | Where to go |
1. Aura AI Chat — invitation-only: a business question and a follow-up | |
2. Evaluate applications: encrypted databases, messaging or private inference | |
3. Learn and build through MCP from Cursor, Claude Desktop or VS Code |
Understand the differences, keys and evidence. Start with the AI application for the product experience; use MCP to explore the supported numeric integration. AI Chat is invitation-only. Request an invitation or arranged demonstration via gen@afhe.io. The public MCP learning demo is open and does not require a Chat invitation.
MCP helps users learn how FHE works, integrate supported encrypted calculations into applications, and explore their potential. Start with a guided public lesson, then map your own use case to its computation, keys and result workflow. The public demo needs installation and host configuration. Confidential application integration also needs local encryption, key custody and authenticated service setup; browser WASM integration is planned. From learning to building.
The AI application is accessed separately; this package does not run model inference. Its executable tools provide encrypted numeric computation. Aura's confirmed FHE database and FHE-AI LLM inference applications are completed and available on request via gen@afhe.io. 中文说明.
These buttons open instructions. Install Node.js 20+, then run these two commands in Terminal (macOS/Linux) or PowerShell (Windows):
npm install -g @aurafhe/mcp@preview
aura-fhe-mcp --config cursor --demoUse claude or vscode instead of cursor for those apps. Paste the generated
settings into your app's MCP configuration and reconnect.
Exact steps and paste locations. No Git, npm account, local
computation engine or demo key setup is needed. Settings must be generated on the
machine that runs the MCP; paths from a cloud workspace do not work on your laptop.
Then ask:
What can I do with Aura? Show me the AI application first, then explain what I can run here.
Your assistant explains the published AI benchmark and the three experiences:
the separate AI application, application evaluations, and the MCP integration.
Within MCP, choose the optional public-sample lesson or the developer tools.
Connecting alone does not start a chat or
calculation. In clients with MCP prompts, choose What can I do with Aura?
(aura_demo). The overview works without a backend connection; a separate status
check or learning lesson checks the service.
This is release 0.5.0-rc.8, an explicitly labelled preview. To pin it, install
@aurafhe/mcp@0.5.0-rc.8. To update an existing installation, run the installation
command again, regenerate settings and fully restart your client.
Related MCP server: enVector-MCP-Server
Add, subtract, multiply and divide encrypted integers and floats.
Compose sums, products, float averages and weighted sums on the coprocessor.
Discover available operations, measure request latency and save encrypted results.
Explore application access and distinguish published benchmarks from live observations.
Numeric arithmetic is a core FHE use case. Binary-operation support is a planned capability milestone; it is not required for numeric computation to be homomorphic. An application uses these building blocks with its own input, key and output workflow. SQL, retrieval and model inference are not current MCP tools. Tool reference.
Teach me encrypted computation using Aura's public sample lesson.
The assistant calls aura_start with experience: "learn", explains FHE, prepares
encrypted public samples, requests a calculation and saves the encrypted result.
It handles identifiers for you. Ask questions or say “one step at a time.”
The aura_learn MCP prompt starts this same lesson.
The first example combines public totals 25 and 17. Expected sum: 42. The MCP returns encrypted data; the separate synthetic verifier checks actual numerical results outside the conversation. Follow-up samples 7.5, 2.5 and 2 support an encrypted average. Observed checks and reproduction.
Public-sample demo: Aura manages its demo keys and can decrypt its demo data. Use public examples here. Confidential deployments need a verified key-custody and authorization design. Anything entered in your host assistant is visible to that assistant provider. Security model.
Describe your application to the assistant using public or invented information. It can help map your task to available operations and explain the client/server configuration you need. Integration steps and the path to simpler setup.
The package contains the connection adapter, not the proprietary computation engine. It accepts no custom plaintext, secret keys or arbitrary file paths in tool arguments. Operator-provisioned encrypted bundles require an authenticated compute-only deployment and separate recipient processing; this path is not Verified mode. The unpublished reference branch is a separate engineering fixture.
Application walkthrough · Investor demo · Architecture · Verification · Security
Compute on sealed handles (and optional plaintext, which is sealed first). Does not decrypt unless reveal=true. Prefer fhe_private_eval for one-shot AI work.
Inputs: op (required), domain (required), inputs (required), reveal.
Decrypt a handle from fhe_encrypt / fhe_compute / fhe_private_eval. Only call this when the user asked to see the plaintext result.
Inputs: handle (required).
Encrypt a value for private AI compute. Returns a short handle (ct_…) the model can pass to later tools. The model must not treat the handle as plaintext.
Inputs: value (required), domain (required), raw, public.
List the private ops this MCP server can run (add, mean, concat, …). Pass these names to fhe_private_eval.
Inputs: no named parameters.
The main agent tool. Seals inputs, runs the op (add, mul, mean, concat, …) without showing intermediates, and optionally reveals only the final answer. Prefer this over encrypt/compute/decrypt.
Inputs: op (required), domain (required), values (required), reveal.
MCP health check. Call this first so the agent knows the private-compute tools are online.
Inputs: no named parameters.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 6 tools
fhe_status and fhe_ops are clearly distinct, and fhe_encrypt/fhe_decrypt have narrow roles, but fhe_compute and fhe_private_eval overlap substantially. The descriptions partially resolve this by positioning fhe_private_eval as the preferred high-level tool, but an agent could still misselect between low-level compute and one-shot evaluation.
All tools share a consistent fhe_ prefix and snake_case style, which makes the family readable and predictable. The main deviation is that status and ops are nouns rather than verb commands, while private_eval is an adjective-verb rather than a clean verb_noun form.
Six tools is a well-scoped set for a private-compute MCP server. Each tool maps to a necessary step in the health-check, operation-listing, encryption, compute, and decryption workflow, and none feel extraneous.
The server covers the core private-compute lifecycle: health, capability discovery, encryption, computation, high-level evaluation, and decryption. Minor gaps like explicit key management or handle inspection are not required for the described use case, so the surface is functionally complete.
MCP-native Trust Infrastructure for AI Agents. Persistent encrypted memory with Trust Quotient.
A secret store for AI agents: the agent never sees the plaintext.
Sovereign E2E cloud storage for AI agents. Zero-knowledge, RGPD-compliant.
Private, portable memory and reusable skills for AI agents.
