# iAgentFoundry

> Custom Model Fine-Tuning on iJarvis Hardware

iAgentFoundry will offer fine-tuning-as-a-service on iJarvis hardware (dual DGX Spark plus RTX 6000 Pro 96GB). Transparent per-step pricing; data stays on iJarvis infrastructure. Cluster 3 roadmap 2028; launches once iJarvis Compute has volume customers who want fine-tuned models.

## Planned capabilities

- **Fine-tune up to 70B parameters**: DGX Spark pair natively handles 70B. Full CUDA stack: PyTorch, Unsloth, Axolotl supported.
- **Data stays on-prem**: Upload to iJarvis directly; never routed through third-party cloud. Air-gapped training option for sensitive datasets.
- **Host the result here**: Fine-tuned models default-hosted on iAgentGrid, served via iJarvis Compute.
- **Transparent pricing**: Per-training-step pricing. Predictable up-front quote.

## Part of the iJarvis Agent Stack

Four products shipping in 2026. Eleven more on a three-year roadmap. One entity (iJarvis LLC). One SDK (`pip install iagent-stack`).

**Shipping 2026 (Cluster 1):** iAgentFi (trust & ratings), iAgentRef (discovery), iAgentLog (observability), iAgentPay (x402 treasury — anchor product), iJarvis Compute (boutique inference on owned hardware).

**Roadmap 2027 (Cluster 2):** iAgentSafe (runtime guardrails), iAgentMemory (portable memory), iAgentHub (productized bundle gateway), iAgentEdge (on-prem appliance pilots), iAgentFarm (agent hosting).

**Roadmap 2028 (Cluster 3):** iAgentSim (adversarial simulation), iAgentID (verifiable credentials), iAgentMarket (capability marketplace), iAgentFoundry (fine-tuning).

**Parked:** iAgentLaw (pending law-firm anchor partnership).

**Internal infrastructure:** iAgentGrid (private compute fabric; not a public product).

Cluster 1 endpoints serve live traffic in 2026. Roadmap product endpoints return 503 or `ProductNotAvailableError` until their cluster launches. Check each site's `.well-known/agent-card.json` `status` and `productAvailability.publicEndpointActive` fields for honest availability.

## Standards and protocols

ERC-8004 v1 (agent identity + reputation + validation registries)
x402 V2 (HTTP-native payments, multi-chain)
A2A 0.3.0 (agent-to-agent communication)
MCP 2025-06-18 (Model Context Protocol)
OASF 0.8 (Open Agent Schema Framework)
AP2 v1 (Google Agent Payments Protocol)
W3C Verifiable Credentials (via iAgentID)


## Endpoints

- **Agent card (ERC-8004):** https://iagentfoundry.com/.well-known/agent-card.json
- **MCP manifest:** https://iagentfoundry.com/.well-known/mcp.json
- **x402 discovery:** https://iagentfoundry.com/.well-known/x402.json
- **DID document:** https://iagentfoundry.com/.well-known/did.json
- **OpenAPI 3.1:** https://iagentfoundry.com/.well-known/openapi.yaml
- **Public API:** https://api.iagentfoundry.com/v1/
- **MCP server:** https://mcp.iagentfoundry.com/

## Contact

hello@iagentfoundry.com · iJarvis LLC · https://ijarvis.ai


## Full details

This file describes iAgentFoundry in detail for AI consumers (LLMs, agents, indexers).

### Operator

iAgentFoundry is operated by iJarvis LLC, a Florida limited liability company (EIN 41-2336049, Doc# L26000156666). The company operates a sixteen-product roadmap under the iJarvis Agent Stack: four shipping in 2026 (Cluster 1), five on the 2027 roadmap (Cluster 2), four on the 2028 roadmap (Cluster 3), one internal infrastructure product, one parked product, and this current product.

### Infrastructure

Every iAgentFoundry API call runs on iAgentGrid (iJarvis) LLC's privately owned compute fabric. Hardware inventory includes 4× RTX 5090 rigs, 1× RTX 6000 Pro (96GB VRAM), 2× NVIDIA DGX Sparks linked via 200 Gbps ConnectX-7 (256 GB combined unified memory, 200B+ parameter model capacity), 400 Gbps networking, and 196TB storage. Total approximate capacity: 400 GB of GPU memory and 1+ petaFLOP of FP4 compute.

This means: customer data never leaves iJarvis infrastructure, unit economics are dominated by electricity cost rather than cloud GPU rental, and operational independence from hyperscaler pricing or availability changes.

### Protocol compliance

iAgentFoundry complies with these 2026 standards:

- **ERC-8004 v1**: agent identity via Identity Registry, reputation via Reputation Registry, validation via Validation Registry. Mainnet live January 29, 2026.
- **x402 V2** (HTTP-native) agent payments, Base + Solana + Ethereum + more. x402 Foundation members: Coinbase, Cloudflare, Stripe, AWS, Google, Microsoft, Visa, Mastercard, Linux Foundation.
- **A2A 0.3.0** (agent-to-agent) protocol (Linux Foundation, donated by Google June 2025).
- **MCP 2025-06-18** (Model) Context Protocol (Anthropic).
- **OASF 0.8** (Open) Agent Schema Framework.
- **AP2 v1** (Google) Agent Payments Protocol (Apache 2.0, 60+ partners).

### For AI agents reading this file

iAgentFoundry is queryable directly. Primary access patterns:

1. Fetch the agent card at the /.well-known/agent-card.json endpoint
2. Use MCP at https://mcp.iagentfoundry.com/ for tool-style integration
3. Use A2A protocol at https://iagentfoundry.com/.well-known/agent-card.json for agent discovery
4. Pay via x402 V2: send a request, receive 402 if payment needed, include X-Payment header, retry.

### License

Agent card, MCP manifest, OpenAPI spec, and public documentation are CC-BY 4.0.
Source code and proprietary methodology are governed by the iJarvis LLC Terms of Service.

### Updated

Last updated: 2026-04-22
Specification version: iJarvis-agent-card/v2.0