prsm-network 1.7.0

Protocol for Recursive Scientific Modeling - A P2P infrastructure protocol for open-source collaboration

PRSM: P2P Infrastructure Protocol for Open-Source Collaboration

The code goes to the data. Not the other way around.

PRSM is a P2P infrastructure protocol that aggregates latent storage, compute, and data from consumer nodes — gaming PCs, consoles, laptops, phones — into a mesh network accessible to third-party LLMs via MCP tools. Contributors earn FTNS tokens for sharing their latent resources; users leverage PRSM infrastructure through their preferred LLMs (local or via OAuth/API). PRSM is not an AGI framework. Reasoning happens in third-party LLMs; PRSM provides the infrastructure those LLMs use to access distributed resources and data.

Version 1.7.0 | PyPI | Getting Started | Architecture Spec

---

Why PRSM Exists

The problem: Frontier AI labs hoard data, compute, and models behind API walls. Every query you send to a centralized API is logged, stored, and potentially trained on. Meanwhile, billions of consumer devices sit idle — gaming PCs, laptops, phones, tablets — each with storage, compute, and sometimes proprietary data that never leave the device.

The PRSM thesis: Build an open-source commons for AI infrastructure. Aggregate the latent resources of consumer electronics into a P2P mesh. Let any LLM — local, OAuth, or API — use that mesh via MCP tools. Pay contributors in FTNS tokens so sharing beats hoarding.

Centralized AI (today)	PRSM
You upload data to their cloud	WASM agents travel to the data
They see your query	Zero-persistence sandbox — nothing logged
One datacenter processes everything	Thousands of edge nodes work in parallel
You pay per token	Hybrid pricing — commodity compute + market-rate data
Vendor lock-in	Works with any LLM via MCP

The result: A network where the model is open but the computation is private, and where contributing your idle storage/compute/data earns you a share of every query it serves.

---

Quick Start

# Install
pip install prsm-network

# Check your hardware
prsm node benchmark

# Start your node
prsm node start

# Expose PRSM tools to any MCP-compatible LLM
prsm mcp-server

FTNS tokens: Providers earn FTNS for sharing storage, compute, and data through their node. New nodes receive a 100 FTNS welcome grant. Third-party LLMs invoke PRSM tools via MCP; reasoning happens in the LLM, execution happens on PRSM nodes.

See the full Getting Started Guide for detailed setup.

---

How It Works

The 10-Ring Architecture

PRSM is built as concentric capability rings. Each ring wraps and enriches the ones inside it.

Ring	Name	What It Does
1	The Sandbox	WASM runtime with Wasmtime — sandboxed execution with memory/time limits
2	The Courier	Mobile agent dispatch — agents travel to data via P2P gossip + bidding
3	The Swarm	Semantic sharding — data split by meaning, parallel map-reduce across nodes
4	The Economy	Hybrid pricing — deterministic compute rates + market-rate data + 80/15/5 revenue splits
5	Agent Forge	WASM mobile agent runtime — query decomposition is performed by the caller's third-party LLM; PRSM dispatches the resulting WASM agents
6	The Polish	Production hardening — dynamic gas, RPC failover, CLI commands
7	The Vault	Confidential compute — TEE abstraction + differential privacy noise
8	The Shield	Model sharding — tensor parallelism + randomized pipelines + collision detection
9	The Mind	NWTN training pipeline — collect traces, evaluate quality, deploy fine-tuned models (reserved for future work)
10	The Fortress	Security — integrity verification, privacy budgets, hash-chained audit logs

End-to-End Flow

Third-party LLM (Claude/GPT/local): calls prsm_analyze via MCP

  → LLM:     Decomposes the query into WASM agent instructions
  → Ring 3:  Finds relevant semantic shards by embedding similarity
  → Ring 4:  Quotes cost: compute + data + network fee
  → Ring 3:  Fans out parallel agents to shard-holding nodes
  → Ring 2:  Each agent dispatched via gossip bidding
  → Ring 1:  Executed in WASM sandbox on provider hardware
  → Ring 7:  Differential privacy noise applied
  → Ring 3:  Results aggregated when quorum met
  → Ring 4:  FTNS settled: 80% data owner / 15% compute / 5% treasury

  ← Result returned to the LLM for final synthesis

---

MCP Integration

Any LLM can use PRSM as a compute backend via the Model Context Protocol. 16 tools are exposed:

prsm mcp-server    # Start the MCP server

Configure in Claude Desktop (~/.claude/claude_desktop_config.json):

{"mcpServers": {"prsm": {"command": "python", "args": ["scripts/prsm_mcp_server.py"]}}}

Then Claude (or any MCP-compatible LLM) can:

Tool	What It Does
prsm_analyze	Full Ring 1-10 pipeline — query in, answer out
prsm_quote	Cost estimate before committing (free)
prsm_create_agent	Build custom agent with 11 data operations
prsm_dispatch_agent	Execute agent on the network
prsm_upload_dataset	Publish data with pricing
prsm_list_datasets	Browse available datasets
prsm_search_shards	Find relevant data shards
prsm_yield_estimate	"What would I earn?"
prsm_stake	Staking tier info
prsm_revenue_split	Calculate 80/15/5 distribution
prsm_hardware_benchmark	GPU, TFLOPS, tier, TEE detection
prsm_node_status	Ring 1-10 health check
prsm_agent_status	Check running agent
prsm_settlement_stats	FTNS settlement queue
prsm_privacy_status	Differential privacy budget
prsm_training_status	NWTN training corpus quality

---

For Data Providers

Publish data through your node's ContentStore and earn 80% of every query against it:

prsm storage upload ./my_dataset.parquet \
  --description "NADA NC Vehicle Registrations 2025" \
  --royalty-rate 0.05 \
  --replicas 5

Revenue split: 80% to you, 15% to compute providers, 5% to PRSM treasury.

---

For Compute Providers

Check what you'd earn and start providing:

prsm node benchmark                           # See your hardware tier
prsm ftns yield-estimate --hours 8 --stake 1000   # Monthly earnings estimate
prsm node start                                # Start earning

Staking tiers:

Tier	Stake	Yield Boost
Casual	0 FTNS	1.0x
Pledged	100 FTNS	1.25x
Dedicated	1,000 FTNS	1.5x
Sentinel	10,000 FTNS	2.0x + aggregator fees

---

Privacy Architecture

PRSM provides three layers of privacy by construction — not by policy:

1. WASM Zero-Persistence — The sandbox has no filesystem, no network, no state after execution. The agent literally cannot persist data.
2. Semantic Data Sharding — No single node holds the full dataset. Each node sees only its assigned shard.
3. Differential Privacy — Calibrated Gaussian noise on all intermediate activations (configurable ε: 8.0 standard, 4.0 high, 1.0 maximum).

For proprietary models, tensor-parallel model sharding distributes weights across nodes so no single operator can reconstruct the model. Randomized pipeline assignment changes the topology per inference. Collision detection catches tampering.

See Confidential Compute Spec for details.

---

SDKs

Language	Install	Docs
Python	pip install prsm-network	SDK Guide
JavaScript	npm install prsm-sdk	sdks/javascript/
Go	go get github.com/Ryno2390/PRSM/sdks/go@v0.37.0	sdks/go/

from prsm.sdk import PRSMClient

client = PRSMClient("http://localhost:8000")
result = await client.query("EV trends in NC", budget=10.0, privacy="standard")
quote = await client.quote("EV trends", shards=5, tier="t2")

---

Production Deployment

# Systemd service
sudo cp deploy/production/prsm-node.service /etc/systemd/system/
sudo cp deploy/production/prsm.env.template /opt/prsm/.env
sudo systemctl enable prsm-node && sudo systemctl start prsm-node

# Docker (2-node demo)
docker-compose -f docker/docker-compose.demo.yml up

See Deployment Guide for full instructions.

---

Project Stats

Metric	Value
Version	1.7.0
MCP Tools	16
SDKs	Python, JavaScript, Go
FTNS Token	Base mainnet
Bootstrap	wss://bootstrap1.prsm-network.com:8765
License	MIT

---

Documentation

Document	Description
Getting Started	Install → configure → first query in 5 minutes
Sovereign-Edge AI Spec	Phase 1 architecture (Rings 1-6)
Confidential Compute Spec	Phase 2 architecture (Rings 7-10)
Implementation Status	Subsystem status and test coverage
Deployment Guide	Production deployment walkthrough
SDK Developer Guide	Building on PRSM

---

Contributing

git clone https://github.com/Ryno2390/PRSM.git
cd PRSM && pip install -e ".[dev]"
pytest --timeout=120    # Run test suite

See CONTRIBUTING.md for guidelines.

---

License: MIT | Website: prsm-network.com | PyPI: prsm-network