Co‑own the modelsyou help create.
Foundry is the supply-side protocol for 0G. Pool data, compute, and capital. Co-train an AI model. Own a verifiable, revenue-generating share — minted on mainnet, attributed inside a TEE.
- 0G Storage·data
- 0G Compute·TEE
- 0G Chain·mint
- Agent ID·trust
- SDK·infer
- MCP·agents
The Ingot · a co-owned model on-chain
ingot:0x8e2…f4a · attested
Call a co-owned model.Right now.
The request below routes through the OpenAI-compatible Foundry proxy to a live Ingot on 0G Aristotle mainnet. The inference fee is reserved on-chain via the 0G Compute serving broker, and a portion deposits into the Ingot's RevenueSplitter — claimable by its co-owners.
- Ingot:
0x8e2af4a0…0001 - Adapter: OpenAI-compat · also via Vercel AI SDK or LangChain
- Revenue routed: pro-rata to share-ledger holders, pull-claim on chain
POST /api/v1/chat/completionsReproduce with curl
curl https://foundryprotocol.xyz/api/v1/chat/completions \
-H "content-type: application/json" \
-H "x-foundry-ingot-id: 0x8e2af4a000000000000000000000000000000001" \
-d '{"messages":[{"role":"user","content":"Tum kasso assa?"}]}'The problem
Model creation is locked inside companies.
Training a useful AI model takes three things almost nobody has all of: data, compute, and capital. Today, the people who supply each of those capture none of the upside of what they help create. The ownership instrument doesn’t exist.
Computational linguist
Elena
A 50k-sentence labeled corpus for a low-resource language. Today, it sits on a hard drive.
Capital allocator
Marcus
Believes a legal-clause classifier will be valuable. Can't co-invest in a model's creation.
GPU operator
Aiko
Sells compute at spot price. Captures none of the upside of the models she helps create.
Foundry is the protocol where Elena, Marcus, and Aiko all become co-owners of the same Ingot — proportional to the value each one actually adds. Measured inside a TEE. Minted on-chain.
How it works
A foundry, on-chain. Anyone contributes. Everyone who matters, owns.
Five steps from open call to claimable revenue — each one verifiable on 0G mainnet.
- 01
A Forge opens
Anyone creates a Forge for a model they want trained. Data, compute, and capital are escrowed on-chain.
- 02
Contributions roll in
Smiths contribute datasets to 0G Storage, compute credits, or capital — independent wallets, real value.
- 03
Verifiable attribution
Inside a TEE on 0G Compute, baseline + contribution is measured against a secret holdout. Marginal delta = your share.
- 04
Ownership mints
$FORGE-denominated shares mint to contributors proportional to measured marginal contribution. The cap table is on-chain.
- 05
Inference routes revenue
Any 0G dApp calls the Ingot via the SDK. Payment hits the RevenueSplitter. Owners claim, on-chain, automatically.
TEE attestation · Forge #0042
verifiedMultilingual Translator
ingot:0x8e2…f4a · baseline 21.4 → 38.7 BLEU
enclave
sgx
holdout
sealed
quote
0x7a3b…91d
Marginal Δ by contributor
Forge in Public
No fake counters. Just live mainnet activity.
Updated < 4s after every event
Forges live
0
Ingots minted
0
Total contributions
0
External Smiths
0
Revenue distributed
0.00OG
Indexed live from 0G mainnet events. Sparklines stretch toward the current value until per-day buckets ship.
View full dashboard →Build on Foundry
Three lines of code. Your agent calls a co-owned model.
Foundry is open by design. Drop our SDK into any AI agent project and inference a Foundry Ingot — revenue routes back to the contributors who made it good.
OpenAI-compatible
Drop our endpoint into any OpenAI client. Same shape. Just point the base URL.
Receipts on every call
Each inference returns an attested receipt — TEE quote + on-chain revenue tx.
Adapter-first
Native plugins for Vercel AI SDK, LangChain, and the OpenAI proxy. CLI included.
MCP for agents
npx @foundryprotocol/mcp gives Claude Desktop, Cursor, and Cline a co-owned model as a tool.
"tk-k">import { Foundry } "tk-k">from '@foundryprotocol/sdk'; "tk-k">const foundry = "tk-k">new Foundry({ contracts: 'aristotle' }); "tk-k">const { output, receipt } = "tk-k">await foundry.inference.run( 'ingot:0x8e2…f4a', { input: 'Translate to French: Good morning' });"tk-c">// ▲ revenue routes back to the Ingot's"tk-c">// co-owners on-chain, automatically. "tk-k">console.log(receipt.inferenceTxHash);"tk-k">console.log(receipt.revenueTxHash);Honesty as a feature
Every feature is Real, Stub, or Roadmap. No fourth category.
Real ships on mainnet or passes full tests. Stub is wired end-to-end but waiting on a final 0G primitive — usually one env var away. Roadmap is designed, dated, but unbuilt. This table is the source of truth.
6 Solidity contracts — full integration test coverage
LOO attribution math + on-chain digest commit
$FORGE-denominated proportional ownership minting
RevenueSplitter — pull-payment claims
@foundryprotocol/sdk + Vercel AI / LangChain / OpenAI adapters
Lineage Graph — read directly from Ingot.meta() on-chain
Forge in Public dashboard — counts derived from chain logs
TEE attestation envelope — ECDSA-signed, ecrecover-verifiable
Deploy infrastructure — Makefile + Anvil dry-run in CI
Contracts deployed to 0G Aristotle mainnet
Inference proxy — OpenAI-compatible, routes via 0G Compute broker
AI-assisted Forge wizard — 0G Compute structured generation
TEE-hosted eval coordinator (SGX/TDX quote)
Live revenue tx hash in inference receipts
Shapley / influence-function attribution
Forge governance via $FORGE
Secondary market for Ingot shares
Reputation-weighted contribution caps
Full external audit
Foundry doesn’t use 0G. Foundry grows it.
Join a Forge as a contributor. Spin up your own. Or build on top — your agent project can call any Foundry Ingot in three lines.