claudeway 0.3.0

Verifiable multi-agent consensus for Claude. The coordination layer frameworks punted on.

Claudeway

Verifiable multi-agent consensus for Claude. The coordination layer that frameworks punted on.

Every agent framework can make one agent answer. Most can run several in parallel. None of them make agents genuinely agree — with the disagreement surfaced and the result cryptographically signed. That's Claudeway.

Buzz gave agents a room. Goose gave one agent tools. LangGraph makes you wire coordination by hand. Claudeway is how agents reach agreement.

---

The 30-second pitch

from claudeway import AgentConfig, Swarm, SwarmConfig, Task

swarm = Swarm(SwarmConfig(
    name="ArchReview",
    agents=[
        AgentConfig("StrongConsistency", "Distributed Systems Engineer", "..."),
        AgentConfig("Operations", "SRE / Platform Lead", "..."),
        AgentConfig("Pragmatist", "Staff Engineer", "..."),
    ],
), api_key=...)

result = await swarm.process(Task(
    id="q1",
    description="Active-active Postgres, FoundationDB, or eventual consistency for payments?",
    input_data={},
))

print(result.result["final_answer"])
print(f"agreement: {result.result['agreement']:.0%}  disagreed: {result.result['disagreed']}")

Three specialists. One signed answer. Disagreement surfaced, not averaged away.

Why this exists

The agent-framework field is crowded but lopsided:

Tool	What it solves	The gap
LangGraph ($1.25B)	Production orchestration, manual graphs	Coordination is DIY wiring
CrewAI	DX, fast prototyping	Coordination is shallow; +48% token cost
Goose (Block)	Single-agent harness + MCP	No multi-agent at all
Buzz (Block)	The room agents talk in	Explicitly punted on coordination
Anthropic SDK	Base Claude calls	No coordination layer
MCP marketplaces	Tool distribution	No agreement primitive

The whitespace, plain: everyone built where agents act or where agents talk. Nobody built how agents agree. Claudeway fills exactly that.

What makes it defensible

Consensus results aren't text — they're signed, tamper-evident attestations anyone can verify:

from claudeway import ConsensusReceipt, Ed25519Backend, MLDSABackend

receipt = ConsensusReceipt.from_result(result, swarm_name="ArchReview", task_id="q1")

# Sign with Ed25519 (default — compact, fast, everywhere):
Ed25519Backend().sign_receipt(receipt, ed_private_key)

# Or sign the same payload with ML-DSA-65 (post-quantum, FIPS 204):
MLDSABackend().sign_receipt(receipt, mldsa_private_key)  # receipt.algorithm == "mldsa65"

# Anyone, anywhere, can verify later — tampering with the answer
# invalidates the signature under either scheme.

• Ships with both Ed25519 and ML-DSA-65 backends — switch with one kwarg. Ed25519 for speed and footprint; ML-DSA-65 (NIST level 3, FIPS 204) for attestations that must stay verifiable across the transition to cryptographically relevant quantum hardware
• Same receipt, same canonical payload hash — the SignatureBackend ABC isolates the crypto; consensus code is untouched either way
• Three transports — the same signed receipt renders as plain JSON, a W3C Verifiable Credential, or a Nostr NIP-78 event that drops into a Buzz room

This is the layer that makes consensus worth acquiring rather than reimplementing in a weekend.

Buzz adapter — consensus that lands in a Buzz room

The Nostr transport is exercised end-to-end against a reference relay:

from claudeway.transports import to_nostr_event

event = to_nostr_event(receipt, private_key_hex=nostr_key, d_tag="room-42")
# event.kind == 30078 (NIP-78 addressable)
# event.sig is BIP-340 Schnorr over the sha256 of the NIP-01 serialization
# event.content carries the signed JSON receipt — any Nostr client decodes it

The event Claudeway produces verifies clean under nak (the reference Nostr CLI) and round-trips through nak serve — publish, subscribe, read back, receipt signature still verifies. See TESTLOG.md for the four-layer evidence trail (BIP-340 KAT, 164-test suite, nak verify, end-to-end demo) and examples/buzz_consensus_demo.py for the showcase.

Install

pip install claudeway              # the SDK (4 deps, lean)
pip install claudeway[mcp]         # + expose consensus as an MCP server
pip install claudeway[nostr]       # + sign Nostr events for Buzz interop
pip install claudeway[pq]          # + ML-DSA-65 post-quantum signature backend
pip install claudeway[langgraph]   # + LangGraph StateGraph adapter
pip install claudeway[maf]         # + Microsoft Agent Framework adapter
pip install claudeway[crewai]      # + CrewAI @tool + Flow adapter
pip install claudeway[docs]        # + build the docs site locally

Use it three ways

1. As a Python SDK

from claudeway import Swarm, WeightedVote, Debate

# Cheap default — one round, picks highest-confidence answer
swarm = Swarm(config, consensus=WeightedVote())

# For hard questions — agents see peers' answers and revise once
swarm = Swarm(config, consensus=Debate())

2. As an MCP server (any agent can call consensus)

claudeway-mcp            # stdio — for Claude Code, Cursor
claudeway-mcp --http     # HTTP/SSE — for remote agents

Now any MCP-capable agent (Claude Code, Goose, Buzz rooms) gains two tools: reach_consensus and verify_consensus. One tool call gets a signed agreement — no framework to learn, no graphs to wire.

3. As a coordinator (hierarchical decomposition)

from claudeway import Coordinator, CoordinatorConfig

coord = Coordinator(CoordinatorConfig())
coord.add_sub_agent("Researcher", ...)
coord.add_sub_agent("Analyst", ...)
result = await coord.coordinate(task)
# The coordinator's plan is parsed for real, specialists run in dependency
# order (parallel where independent), results synthesized.

Examples

• examples/quickstart.py — 15 lines, 3 agents → signed receipt
• examples/consensus_demo.py — disagreement surfaced, then resolved via Debate
• examples/coordinator_demo.py — real decomposition + parallel specialists
• examples/buzz_consensus_demo.py — Buzz adapter: 3 agents → signed consensus → live Nostr relay → read back + verified

Features

• Real consensus — WeightedVote (cheap, default) and Debate (revision round when agents disagree, early-exits when they already agree)
• Real decomposition — JSON plan parsing, dependency-respecting parallel execution, specialist routing
• Concurrent — N-agent swarms issue all Claude calls in parallel, not serially
• Verifiable — every result is a signed, tamper-evident receipt
• MCP-native — ships as an MCP server for the 10K+ server / 8M-downloads/month ecosystem
• Claude-native — built on the Anthropic SDK, prompt-caching-friendly
• Lean — 4 base dependencies

Architecture

┌─────────────────────────────────────────────┐
│  Transports (claudeway.transports)          │
│  JSON receipt · W3C VC · Nostr NIP-78 event │
└──────────────────┬──────────────────────────┘
                   │ carries
┌──────────────────▼──────────────────────────┐
│  Signing (claudeway.signing)                │
│  ConsensusReceipt · Ed25519Backend · (PQ)   │  ← the moat
└──────────────────┬──────────────────────────┘
                   │ signs
┌──────────────────▼──────────────────────────┐
│  Consensus (claudeway.consensus)            │
│  WeightedVote · Debate · (your strategy)    │
└──────────────────┬──────────────────────────┘
                   │ aggregates
┌──────────────────▼──────────────────────────┐
│  Core (claudeway.swarm/coordinator/agent)   │
│  Swarm · Coordinator · Runtime · Tools/MCP  │
└─────────────────────────────────────────────┘

See docs/ARCHITECTURE.md for the full design.

Roadmap

Shipped (v0.2.0):

• SDK (Swarm, Coordinator, Agent) with real concurrent execution
• WeightedVote + Debate consensus strategies with cost-guarded early-exit
• Signed receipts with the swappable SignatureBackend ABC and an Ed25519 default backend
• Three transports: JSON receipt, W3C Verifiable Credential, Nostr NIP-78 event
• MCP server (claudeway-mcp) exposing reach_consensus + verify_consensus
• Buzz adapter — verifiable end-to-end against nak serve

Shipped (v0.3.0):

• Post-quantum signature backend — ML-DSA-65 (FIPS 204), pure-Python via dilithium-py. Same receipt, same canonical hash; swap in at the SignatureBackend ABC.
• LangGraph adapter — Claudeway consensus as a single StateGraph node
• Microsoft Agent Framework adapter — Claudeway as a workflow + agent in MAF (AutoGen + Semantic Kernel successor)
• CrewAI adapter — Claudeway as a @tool and Flow; inverts the killer demo so a CrewAI crew can call Claudeway for agreement
• Consensus transparency log — RFC 6962-style append-only log, anchored to Nostr on a cadence
• Streaming consensus events — OnEvent callback surfaces AgentCompleted and ConsensusResolved events as they happen
• Killer benchmark — Claudeway beats single Claude (+7/20) and CrewAI (+5/20) on a hard question, blind-judge scored, Mann-Whitney U
• Adversarial security suite — 20 tests across 7 attack classes + published threat model
• Docs site — jordannewell.github.io/claudeway (mkdocs-material + mkdocstrings)
• Tolerant structured-output parser, substantive <answer> contract
• Stable signing key across adapter runs, async→sync bridge deadlock fix

Next:

• A2A (Agent-to-Agent) protocol adapter — verifiable consensus for Google's agent protocol
• W3C VC presentation exchange flow
• Single-tenant runner — FastAPI + SQLite + dashboard, docker compose up

Deliberately deferred (the "Curtis lesson" — don't build before there's demand):

• Multi-tenancy, billing, Stripe, template marketplace. Re-activated only when paying demand appears. See docs/DEPRECATION.md.

Docs

Full documentation at jordannewell.github.io/claudeway:

• Transports — JSON vs W3C VC vs Nostr NIP-78, and when to pick each
• Adapters — Buzz and LangGraph integration paths
• Benchmarks — Claudeway vs single Claude vs CrewAI
• API reference — auto-generated from the source
• Changelog

The docs site deploys automatically on push to main via the docs.yml workflow. Build locally:

pip install -e ".[docs]"
mkdocs serve          # http://127.0.0.1:8000

Contributing

• CONTRIBUTING.md — branch model, commit rules, test commands, how to add docs
• SECURITY.md — private vulnerability reporting process and SLA
• CODE_OF_CONDUCT.md — Contributor Covenant 2.1
• Bug reports and feature requests: use the issue templates

Development

pip install -e ".[mcp,nostr,pq,dev]"
pytest tests/ -v          # 164 tests (set CLAUDEWAY_TEST_RELAY=ws://localhost:10547 to exercise the relay integration test)
ruff check claudeway/ tests/ examples/

License

MIT