LOG_VER: 0.3.1-ALPHA //
CLASSIFICATION: INTERNAL

AIOS PROTOCOL // AGENT-NATIVE OS // THERMAL-AWARE SCHEDULING

BMETAL

THE OS THAT UNDERSTANDS AI.
AGENTS RUN LONGER. ACT FASTER. STAY SAFE.

SUSTAINED_INFERENCE|30x FASTER TOOL CALLS|CAPABILITY-SCOPED SECURITY

01 / PROTOCOL

WHAT IS
BMETAL?

BMetal is an AI-native operating system that manages AI agents as first-class workloads. Thermal-aware scheduling keeps inference running at full speed for hours — not just minutes. Tool calls execute 30x faster through kernel-routed MCP. Every agent action is capability-scoped and sandboxed at the kernel level.

No cloud dependency. No trust assumptions. No thermal throttling. Your hardware, running at its full potential.

15-40%
SUSTAINED THROUGHPUT GAIN
30x
FASTER TOOL CALLS
5
SECURITY LAYERS
A
SECURITY GRADE

02 / KERNEL

THE STACK

01
THERMAL_SCHED
Thermal-aware inference scheduling. Polls hwmon every 500ms, proactive batch reduction at 70°C — sustains throughput where stock Linux cliff-drops.
thermal_sched.ko
02
VRAM_SCHED
VRAM + UMA memory management. Per-agent quotas, model pinning, zero-eviction under pressure. AMD ROCm, Tenstorrent, ARM64 UMA.
vram_sched.ko
03
ML_CORE
SIMD detection (x86 AVX2/AVX-512, ARM64 NEON), memory pool allocator, zero-copy inference primitives. 11 KUnit tests.
ml_core.ko
04
BMETAL_SEC
Landlock filesystem & network sandbox, 5 security profiles (Minimal → Unrestricted), 12 ioctls, SHA-256 audit hash chain.
bmetal_sec.ko
05
BMETALD
Rust daemon. JSON-RPC over Unix socket. Auth handshake, capability tokens, MCP Bus, 5-tier LanceDB hybrid memory.
bmetald.service
06
INFERENCE
llama.cpp FFI — real streaming inference, greedy/top-k/top-p/temp sampler chain. Stub feature keeps CI fast.
bmetal-inference
07
MEMORY
5-tier: Hot RAM → Warm SSD → Cold zstd → Vault TPM → Archive. FTS5 + LanceDB via Reciprocal Rank Fusion.
bmetal-memory

03 / FEDERATION

MFQP
PROTOCOL

The Metalogue Federated Query Protocol — a cryptographic network for AI-to-AI collaboration without exposing raw prompts.

  • →
    GHOST QUERIESIntent vectors only. TF-IDF + FNV-1a. Your prompt never leaves your node.
  • →
    Ed25519 mTLSPer-node X.509, fingerprint pinning, TOFU feature-gated out of release builds.
  • →
    BILLING LEDGERAppend-only JSONL. Millicredit metering. Compute-time billing in Phase 4.
  • →
    VEC2VECCross-model coalition querying. Translate semantic intent across incompatible embedding spaces.
YOUR NODEBMETALDEd25519 CERTPEER-AGHOST QUERYPEER-BGHOST QUERYPEER-CGHOST QUERYPEER-DGHOST QUERYBILLINGLEDGER.JSONL

04 / BUILD STATUS

ALPHA
0.3.1

KERNEL MODULES● SHIPPED
REAL INFERENCE (llama.cpp FFI)● SHIPPED
HYBRID MEMORY● SHIPPED
SECURITY GRADE: A● SHIPPED
MFQP FEDERATION v1.0● SHIPPED
SHELL + TUI● SHIPPED
PYTHON / RUST / C SDK● SHIPPED
1,008+ TESTS PASSING● SHIPPED
TPM KEY STORAGE◌ PHASE 4
TT-METAL BACKEND (Phase 4)◌ PHASE 4
THIRD-PARTY AUDIT (Phase 4)◌ PHASE 4
FILE UNDER: ADRENALIN RUSH
GET THE ISO
x86-64 / ARM64 / AMD ROCm
SHA256 VERIFIED · GPG SIGNED

05 / DOCS + SDK

START
BUILDING

01
QUICKSTART
Boot BMetal ISO, register your first agent, run inference in 5 minutes.
→
02
PYTHON SDK
pip install bmetal-sdk. Agent.ask(), run(), remember(), recall(). 84 tests.
→
03
RUST SDK
Full async API, SyncAgent wrapper, 38 tests. cargo add bmetal-sdk.
→
04
MFQP PROTOCOL
Ghost query format, mTLS bootstrap, CLI reference, threat model.
→
05
KERNEL MODULE DEV
Writing LKMs for BMetal. vram_sched API, KUnit testing, ARM64 cross-compilation.
→
06
MCP SERVERS
Build and publish custom MCP servers. Template with 30-min README, SHA-256 verify.
→