Phase III · Autonomous Agentic Control Tier

Claude, wired into a real GPU pipeline over stdio.

The MCP-native control plane. omnipulse-agent parses operator requests in natural language and routes them to the Rust orchestrator over line-delimited JSON-RPC 2.0: without ever touching tensor data.

Operator prompt: Check if this 30-second drum loop is a derivative of anything in our licensed catalogue.

Idle·

LLM parses prompt·

RPC: generate_fingerprint·

RPC: compare_fingerprints·

Done

Pane 1: Anthropic Claude

Pane 2: JSON-RPC bus

No frames yet.

Pane 3: Rust + GPU kernel

shm_open(/a3f7b2c1d4e5f6a7b8c9d0e1)

mmap(read-only, 5292000 bytes)

spawn_blocking(kernel.run)

cxx::run_wst_pipeline(ptr=0xDEAD0000, len=1323000)

WSTEngine<HopperTag, 8, 16>::forward()

HNSW insert(fp_01J4, dist=SW₁)

What you get

→omnipulse-agent on PyPI: the Python control plane (Anthropic SDK + MCPClient + SharedMemoryManager).
→Out-of-the-box compatibility with Claude Desktop, Cursor, and any other MCP host.
→End-to-end audit trail: every tool call is a JSON line you can replay.

What it solves

Most 'AI compliance' products bolt an LLM on top of an opaque pipeline. OmniPulse inverts it.
The LLM is a router, not a decision-maker. The deterministic GPU pipeline produces the legally-significant fingerprint.

Pricing

AGPL-3.0: free for research & open source

Enterprise use cases

In-house compliance copilot

A single Claude instance that triages incoming media, calls generate_fingerprint, then compare_fingerprints, and surfaces the match probability.

Forensic replay

Every JSON-RPC frame is appended to a write-once log: the audit story writes itself.

MCP-marketplace deployment

Publish the toolset to any MCP host and your team gets an agent that cannot fabricate fingerprints because the schema rejects free-form parameters.

Quickstart

pip install omnipulse-agent omni-wst-core
cargo install omnipulse-mcp
export ANTHROPIC_API_KEY=sk-ant-...
python -m omnipulse_agent.run --wav your.wav

↗ View on PyPI Star omnipulse-agent on GitHub ↗Talk to founders →