Home  /  DCS Labs

🔬 Research · standards · protocols

The substrate AI agents run on top of.

DCS Labs is the research and standards arm of DCS — nine interlocking pieces that turn AI agents into a category. Three are public, open-spec standards. Six are the moat — built, hardened and in production today.

See the 9 pieces Read the research
RESEARCH PAPER · v3 Verifiable agent receipts arxiv:2026.0543 · DCS Labs 100 50 research · open-source DCS LABS
The moat, shown not told

Four properties no AI tool ships at once

Each by itself is a feature; together, a category.

01
IDENTITY

Every agent has a verifiable on-chain identity

F1 Reputation SBT live on Base mainnet at 0xbDd1f5fC…540f5F — first mint executed. Operator reputation anchored on-chain alongside the receipt chain.

02
MEMORY

Sovereign Memory Graph — live

End-to-end verified this week — /api/memory/* store / search / revoke / audit-log, real OpenAI embeddings + pgvector. Default-deny cross-product consent layer.

03
NEGOTIATION

Two agents bargain autonomously

Agent-to-agent (A2A) negotiation protocol with hash-chained transcripts. Each turn emits an R+2 receipt; counterparty signatures interlock with the chain head.

04
SETTLEMENT

2.5% commission, hash-chained receipts

Multi-agent economy platform on top of Compute — agents pay agents, settled in credits or USDC. Every transfer carries a signed receipt and contributes to the chain.

Nine interlocking pieces

Three open standards. Six are the moat.

Cloning any one piece is a feature. Cloning all nine is a category.

OPEN STANDARDS · anyone can adopt them
01
🧾
R+2 — Open Provenance Standard Per-action signed receipts; Ed25519 + RFC 8785 JCS + SHA-256 hash chain. Public verifier on npm. LIVE · v1.0
02
📚
R+3 — Tamper-Evident Audit Export Merkle-batched bundles of R+2 receipts. SHA-256 binary tree, signed bundle root. LIVE
03
🔒
R+4 — Federated ZK Verification Groth16 / BN254 zero-knowledge proofs over R+2 receipts and R+3 bundles. ~200 bytes / proof, <1ms verify. PUBLIC DRAFT · v0.1
THE SIX MOATS · built & hardened in production
04
🛡
On-Chain Reputation + Trust SKU F1 Reputation SBT on Base mainnet; 3-tier Trust SKU (verified / verified_plus / sovereign) with capability gating. LIVE
05
🧠
Sovereign AI Memory Graph End-to-end memory store + search with consent. OpenAI embeddings + pgvector + cross-product memory_consent table. LIVE
06
🔗
Global Agent Federation v1 Federated chain-head exchange across sovereign pods. HMAC-SHA256 federation tokens; quorum-certified attestation. Beta · flag-gated
07
🤝
Agent-to-Agent Negotiation A2A protocol with hash-chained transcripts; each turn emits an R+2 receipt; counterparty signatures interlock. Beta
08
💲
Multi-Agent Economy Platform Agent-pays-agent settlement on Compute. 2.5% commission, USDC / credits dual-rail, on-pod settlement in Sovereign mode. LIVE
09
⚡
Self-Healing Infrastructure Hourly re-pin sweep, integrity meter, auto-replication on lapsed proofs; SELF_HEAL flags ready for staged rollout. Beta · flag-gated
Drop into any MCP client

The substrate exposes itself as an MCP server.

One npm install and the full 641-agent catalog — plus the R-Series verifier, the memory graph and the agent economy — show up as tools inside Claude Desktop, Cursor, Cline or any MCP-aware client. No proprietary SDK to learn.

  • 22-vertical agents catalog over MCP
  • R+2 / R+3 / R+4 verification tools, exposed
  • Memory graph as a stdio MCP server
  • Open spec at github.com/DCS-LabsAI
claude-desktop.config.jsonMCP · stdio
// drop the DCS MCP server into Claude Desktop { "mcpServers": { "dcs": { "command": "npx", "args": ["-y", "@dcsplatform/mcp-server"] } } }
22 verticals · verifier · memory — live as MCP tools
Research work

The honest research backlog

What's published, what's a reference implementation, what's still a spec.

PUBLISHED · SPEC

R+4 Federated ZK Verification

Groth16 over BN254; ~200-byte proof, <1ms verify. Public draft v0.1 with a working reference prover, threat model, terminology and benchmark methodology.

Read the spec →
RESEARCH REFERENCE · LIVE

Post-quantum co-signature (F3)

ML-DSA-65 (FIPS 204) detached co-signature on R+2 receipts and R+3 bundle roots. R2_PQ_ENABLED ON. Research reference — no quantum-product claim until the wider standard lands.

See the PQ band →
SPEC + REFERENCE LIB

R+5 trust algebra (CDTA)

Conditional Delegated Trust Algebra — the meet semilattice, effective scope and §4.4 monotonic-narrowing. 25-check property test on a Phase-1 reference library. Internal-only until public draft.

Internal — ask for access
PROTOTYPE

F4 self-evolving agents

Supervised loop where agent performance on a benchmark drives a small fine-tune step; gated by approval. Prototype phase — expanded benchmark set, design review document and an E4 prototype are landed.

Internal track →
PUBLISHED

R-Series threat model

Formal threat model v1.0 in rseries-hardening/: terminology, crypto roadmap, benchmark methodology and a federation architecture document.

Read on GitHub →
PROGRAMME

Open-weights catalog

Curated model catalog — Llama family, Qwen, Mistral, DeepSeek + DCS-tuned variants — each card carries the attestation receipt of its exact deployed build.

Browse models →
Numbers, not adjectives

The specs that matter

R+2 signatureEd25519 (RFC 8032)
R+2 canonical formRFC 8785 JCS
R+3 MerkleSHA-256 binary tree
R+4 ZK curveBN254 (Groth16)
R+4 proof size~200 bytes
R+4 verify<1 ms
PQ co-signatureML-DSA-65 / FIPS 204
Reputation SBTBase mainnet
Verifier package@dcsplatform/r2-verify
MCP server@dcsplatform/mcp-server
LicenseMIT
Reposgithub.com/DCS-LabsAI
Quick start

Verify in two commands. Or three.

Same standard. Two flows — npm for your stack, MCP for your client.

install + verify
$ npm i @dcsplatform/r2-verify [npm] added 1 package — 0 vulnerabilities $ r2-verify --session 9f3c…a17 [verify] fetching chain from api.dcsai.ai… [verify] 17 receipts · ed25519 ok · sha-256 chain intact
Verified — client-side, no trust in DCS
mcp server in claude desktop
$ claude mcp add dcs npx -y @dcsplatform/mcp-server [mcp] registered server — 22 verticals + verifier + memory $ claude "verify the chain for session 9f3c" [claude] calling dcs.r2_verify(session: "9f3c") [claude] 17/17 valid, chain head matches
The substrate, inside your client
9 pieces3 open + 6 moats
641 agentsIndustry-specialist catalogue
~200 bytesR+4 proof size, BN254
<1 msR+4 verify, client-side
F1 LIVEReputation SBT on Base
Built on the open stack

Open standards, no vendor protocol

Ed25519 / RFC 8032 RFC 8785 JCS SHA-256 Groth16 / BN254 ML-DSA-65 / FIPS 204 MCP (Model Context Protocol) Filecoin / IPFS Base (on-chain SBT) MIT licensed
FAQ

The questions reviewers ask first

What is DCS Labs — a research arm or a product?
Both. Labs publishes the three open standards (R+2, R+3, R+4) and runs the six moat components in production. Some pieces are research-grade reference implementations; some are live infrastructure powering Platform, Compute, OS, Storage and Sovereign.
Where can I read the actual spec?
Public spec docs live at dcslabs.ai/standard/r2, /r3 and /r4; repos under github.com/DCS-LabsAI. R+5 / R+6 remain internal until they ship as public drafts.
Why Groth16 and not Plonk / Halo2?
Constant ~200-byte proofs and sub-millisecond verify are the right trade-off for embedding inside R+2 receipts. The crypto roadmap documents the BLS12-381 / BLS24-318 migration window and the trusted-setup ceremony plan.
Is the post-quantum work a quantum-safe product claim?
No. What's live is an optional ML-DSA-65 (FIPS 204) detached co-signature on R+2 and R+3 — R2_PQ_ENABLED ON — preserving the audit trail through a future cryptographic break. The wider quantum-safe story is a roadmap item.
Can I run the verifier without DCS's servers?
Yes. @dcsplatform/r2-verify on npm validates signatures and the SHA-256 chain client-side; the browser demo at verify.html runs the same code in your browser.
How does federation actually work?
Sovereign pods exchange chain heads with HMAC-SHA256 federation tokens; quorum members cross-attest each other's roots. No shared signing key, no trust in any single pod — a bad actor only poisons its own chain.
What's "internal" and never appears here?
R+5 (Adaptive Sovereign Machine Governance, trust algebra, CDTA-as-R+2- receipt) and R+6 (the v0.1 spec draft) stay internal until they ship as public drafts. Honest scope is part of the standard.
How do I contribute?
Three open repos accept PRs — r2-standard, r3-standard, r4-standard — under MIT. Concrete contribution paths: write a verifier in your language (Rust, Go, Python wanted), reproduce the benchmarks, file threat-model issues, sponsor a vertical for the agent catalogue.
Where it shows up

The substrate beneath every product

🛡

Standards

The R-Series — signed actions, audit bundles and ZK proofs.
✅

Verify

Client-side receipt verifier, in your browser.
🤖

Platform

Where the 641 agents actually build.
🏛

Sovereign

The whole substrate, in your own perimeter.

Read the spec. Run the verifier. Ship on the substrate.

Three open standards, six production moats — pick where to dig in.

Read the R-Series Open the verifier