Open protocol prototype for AI agents to quote, authorize, execute, verify, settle, and dispute service transactions. Not a payment rail — the semantic layer around agent-to-agent work.
Agent-to-agent communication frameworks (MCP, A2A) enable agents to discover and invoke each other. Payment rails move value. But there's no standard for what happens in between: quoting, authorizing spend, submitting tasks, verifying results, generating receipts, and resolving disputes.
Every team inventing their own clearing logic creates fragmentation, incompatibility, and trust failures at scale.
NeuralClear defines the standardized objects and HTTP endpoints that handle the full transaction lifecycle — from discovery and quoting through mandate authorization, task execution, proof verification, settlement receipt, and dispute resolution.
It does not issue value, custody funds, or replace your payment rails. It defines the commercial semantics and clearing record layer that makes agent transactions verifiable and composable.
Every agent service transaction follows the same lifecycle. NeuralClear standardizes each stage with typed objects and verifiable receipts.
The AI agent stack has emerging standards at every layer. NeuralClear defines the transaction semantics that sits between communication and settlement.
Protocols for agents to discover capabilities, exchange messages, and invoke services. Define how agents talk.
Defines the commercial semantics: how agents quote, authorize, execute, prove, settle, and dispute service transactions. Clearing records for every transaction.
Payment rails that move value between parties after clearing. Define how agents pay. NeuralClear is payment-rail agnostic.
NeuralClear is a protocol prototype under active development. Production deployment is not yet recommended.
Get a NeuralClear server running locally in under five minutes, register an agent, and run your first clearing transaction.
Clone the repository and install with the HTTP server extras:
# Clone the repository git clone https://github.com/Zihuatanejo63/neuralclear cd neuralclear # Install with HTTP server support pip install -e ".[http]"
Start the reference HTTP server on port 8000. SQLite is used for local persistence.
uvicorn server.app:app --port 8000 # Server starts at http://localhost:8000 # Dashboard at http://localhost:8000/dashboard
Use the CLI to register an agent with capabilities and pricing manifest.
# List registered agents python3 -m client.cli agents list # Register a new agent (edit your manifest.json first) python3 -m client.cli agents register manifest.json
Request a priced quote from a registered agent for a specific capability.
# Request a quote from the PDF summarizer agent python3 -m client.cli quote request agent.pdf_summarizer summarize.pdf # Returns a Quote object with quote_id, price, expiry
Submit a TaskRequest against an accepted quote. The agent executes and returns a TaskResult with a Proof.
# Submit a task using an accepted quote ID python3 -m client.cli task submit quote_xxx \ --text "Summarize this PDF text." # Task executes → Proof generated → SettlementReceipt created
# View ledger balances python3 -m client.cli balances # View a specific settlement receipt python3 -m client.cli receipts list # Open dashboard for full UI view open http://localhost:8000/dashboard
NeuralClear defines nine typed objects that cover the full lifecycle of an agent service transaction. All objects are JSON-serializable and version-tagged.
{
"quote_id": "quote_a3f9e2b1",
"spec_version": "0.3",
"provider_agent_id": "agent.pdf_summarizer",
"requester_agent_id": "agent.orchestrator_01",
"capability_id": "summarize_pdf",
"price": 0.05,
"currency": "USD",
"input_hash": "sha256:a1b2c3...",
"expires_at": "2024-01-15T12:05:00Z",
"quoted_at": "2024-01-15T12:00:00Z",
"status": "pending_acceptance"
}
The local sandbox gives you a full NeuralClear environment: HTTP server, SQLite ledger, CLI client, and web dashboard. All running locally with no external dependencies.
Clone the repo and install Python dependencies with HTTP extras.
pip install -e ".[http]"
Runs the FastAPI reference server with SQLite persistence on port 8000.
uvicorn server.app:app --port 8000
python3 -m client.cli agents list
python3 -m client.cli quote request \ agent.pdf_summarizer summarize.pdf
python3 -m client.cli task submit quote_xxx \
--text "Summarize this PDF text."
python3 -m client.cli balances python3 -m client.cli receipts list
The local sandbox ships with three reference agents pre-registered. Each exposes its capabilities, pricing schema, and endpoint via the AgentManifest protocol object.
{
"agent_id": "agent.pdf_summarizer",
"spec_version": "0.3",
"name": "PDF Summarizer Agent",
"description": "Summarizes PDF documents via extractive + abstractive pipeline",
"endpoint": "http://localhost:8001",
"capabilities": [
{
"capability_id": "summarize_pdf",
"name": "Summarize PDF",
"pricing": {
"model": "per_page",
"unit_price": 0.005,
"currency": "USD"
},
"timeout_ms": 30000
}
]
}
NeuralClear is developed in public as an open protocol prototype. Each milestone adds a production-relevant layer to the clearing stack.
NeuralClear is an open clearing protocol. The reference implementation is free and open source. Hosted infrastructure and enterprise support are coming.