Air-gapped AI developer platform

ITAR-compliant AI coding.
On your hardware.

Tetryx deploys as a containerized software appliance on local GPU servers. Open-weights models serve sub-100ms tab completions to VS Code, CLion, and Neovim. No data leaves the building.

CMMC Level 2/3VS Code · CLion · NeovimEmbedded C++ · Rust

Request a Demo

tetryx statuslocalhost

● model qwen2.5-coder-7b-instruct

● engine vLLM 0.6.4 · localhost:8080

● vram 7.2 GB / 24 GB

● latency p50 82 ms

● requests 1,847 today

■ egress blocked 0 attempts

Without Tetryx

Use public cloud AI → ITAR violation

Sending proprietary flight software or ITAR-controlled technical data to OpenAI, GitHub Copilot, or any public cloud API violates export-control law. Penalties are not a line item you can budget for.

With Tetryx

Full AI velocity. Zero egress.

Open-weights models run entirely on your hardware, inside your network boundary. Every completion request is logged for CMMC audit. Nothing crosses the perimeter.

01 / IDE integration

Works inside the tools your engineers already use.

Native extensions for VS Code, JetBrains CLion, and Neovim wire Tetryx into existing workflows. Inline tab completions appear in under 100ms. The local API gateway is OpenAI-compatible, so terminal agents and scripts that already target the OpenAI API work without modification.

—VS Code extension — IntelliSense-compatible inline completions

—JetBrains CLion plugin — embedded C++ and CMake project awareness

—Neovim LSP bridge — via local OpenAI-compatible API endpoint

—Terminal agents — any tool targeting the OpenAI API works unchanged

flight_computer.cpptetryx inline

// Attitude determination and control

void ADCS::propagate(double dt) {

quaternion_.integrate(omega_, dt);

// ↓ Tetryx suggestion (Tab to accept)

J2Perturbation::apply(orbit_, dt);

dragModel_.update(orbit_.altitude(), dt);

}

82 ms·local inference·no egress

$ vllm serve --model /models/qwen2.5-coder-7b

INFO Loading model from /models/qwen2.5-coder-7b

INFO GPU memory allocated: 7.2 GB

INFO Serving on http://localhost:8080

# Completion requests from VS Code extension

POST /v1/completions 200 OK 78ms

POST /v1/completions 200 OK 84ms

POST /v1/completions 200 OK 91ms

# Network: isolated. Zero outbound requests.

02 / Open-source inference

High-throughput open-weights models. No API costs. No rate limits.

Tetryx serves Qwen Coder, DeepSeek, and Gemma models via vLLM and SGLang on your on-premise GPU server. Sub-100ms p50 completions at team scale, with no per-token billing and no external dependency that can be revoked.

—Qwen2.5-Coder, DeepSeek-Coder-V2, Gemma 2 — no vendor lock-in

—vLLM and SGLang engines for maximum GPU throughput

—Continuous vector indexing of local C++ codebases and hardware schemas

03 / Security control plane

Audit-ready from the first completion request.

An embedded security proxy captures full audit logs and enforces data-egress policies. Every completion request is recorded with user, timestamp, model, and disposition. CMMC Level 2 and Level 3 audit reviews start with an export — not a reconstruction.

—Immutable audit log of every completion and egress attempt

—Network policy enforcement — outbound AI API calls blocked at the proxy

—CMMC Level 2/3 and ITAR readiness documentation included

tetryx audit --tailcmmc-l2

14:32:07Z ALLOW completion user=m.chen model=qwen2.5-coder

14:32:11Z ALLOW completion user=j.park model=qwen2.5-coder

14:33:42Z BLOCK egress dst=api.openai.com rule=cmmc-l2

14:33:43Z ALLOW completion user=m.chen model=qwen2.5-coder

14:35:02Z ALLOW completion user=a.vasquez model=qwen2.5-coder

period: 3m 02s · completions: 4 · egress blocked: 1

Licensing

ModelPriceScope

Per-seat annual$1,500 – $2,500 / seatIndividual developer license

Site license$50,000 – $150,000 / yrUnlimited seats, one facility

Hardware is purchased directly through OEM Value-Added Resellers. Tetryx is a pure-software license.

Early access

Access is limited to active design partners.

Tetryx is currently deployed with a small number of venture-backed space and defense engineering teams. Sales cycles run 2–4 weeks on unclassified, ITAR-bound development networks. If your team evaluates AI coding tools under regulatory constraints, we would like to talk.