Two SDKs · One Platform · Zero Secrets Shipped

Build agents. Ship nothing they can steal.

Claude Code leaked 512,000 lines because of a missing .npmignore. MHVR has no npm package. The distribution artifact is a compiled binary and a manifest of hashes. Source never leaves your server. Secrets never touch inference. Prompt injection is structurally impossible.

Security SDK

GentlyOS MHVR

Protect Claude Code and every agentic pipeline you run. LLMs emit hashes, not instructions. The agent's action space is compiled closed at build time.

Prompt injection — structurally impossible
97% token reduction (~60 tokens vs ~2000)
AES-256-GCM encrypted vault — $REF only
Orc namespace permissions — no runtime check
No npm. No dependency chain. Compiled Rust.
4 commands. 30 seconds to secure.

gently init →

Agent SDK

Trollz Deploy

Deploy AI agents to every surface. Embed anywhere. Bind to Discord, Telegram, SMS, Twitch, and the open web. One script tag to living agent.

One-liner embed — <script src="trollz.fun/embed">
Discord / Telegram / SMS / Twitch bindings
.bb neural binary format — model IS the program
SYNTH token economy — agents earn from conversations
Freeze Protocol — competitive social mechanics
WebGL SDF character renderer

trollz.fun →

By the numbers

Not benchmarks. Structural properties of the architecture.

97%

Token reduction
per agent request

Lines of source
shipped to clients

~60

Tokens in LLM
context block

30s

Setup to fully
secured agent

How MHVR protects your agents

Scan

Walk your codebase. Hash every exported function. Output a manifest — hashes and purpose strings. Zero source code leaves.

$ gently scan ✓ Found 47 functions ✓ 3 orc namespaces ✓ Manifest written

Compile

Manifest compiles to a perfect hash map. The agent's entire action space becomes a closed, immutable set. Unknown hash → None → stop.

build.rs → phf::Map "sha256:d4e5" → auth::login "sha256:a1b2" → ui::dashboard "anything_else" → None

Serve

LLM receives ~60 tokens of context. Emits a hash. Hash resolves or doesn't. That's the entire security model. English ≠ 64 hex characters.

POST /grind → "sha256:d4e5..." → dispatch → auth::login → result

vs. everything else

Attack Surface	LangChain / AutoGen / CrewAI	MHVR
Prompt injection	Prompt engineering (prayers)	Hash lookup — structurally impossible ✓
Secret exfiltration	Hope you configured .env right	$REF only — value never in context ✓
Supply chain	npm/PyPI with 500+ dependencies	Compiled Rust binary — no deps to poison ✓
Unbounded action	Runtime guardrails (bypassable)	Compile-time closed action space ✓
Token overhead	~2,000 tokens / request	~60 tokens / request (97% reduction) ✓
Source exposure	Shipped to registry / cloud	Never leaves your server ✓
Setup time	Hours of config + platform accounts	30 seconds, 4 commands ✓

Ship sovereign agents

Security team signs off in one meeting. Finance signs off because token costs dropped 97%.
The protocol spec is open. The implementation is the product.


$ curl -sSf https://gentlyos.io/install | bash
$ gently init
$ gently scan
$ gently serve
# done. agent secured.

Security SDK → Agent SDK → Protocol Spec →