FAQ¶
The questions a careful developer asks before adopting Tractum. Honest answers — including where the SDK is unfinished.
Does this work with non-Claude models?¶
No. Tractum is Claude-specific (per the name). It wraps the Anthropic SDK directly. Adding other model families would require an abstraction layer over the model client; not on the roadmap.
You can swap between Claude model tiers (Haiku, Sonnet, Opus) per agent via AgentConfig.model — useful for cost tuning.
Is this a blockchain? A crypto project?¶
No. Tractum uses standard Ed25519 signatures and Nostr's pub/sub protocol for transport. There is no ledger, no consensus algorithm between nodes, no token, no chain.
The signatures make receipts tamper-evident. Nostr makes them globally publishable. That's the entire "crypto" surface — both are pre-blockchain cryptography from the 1990s (Ed25519) and 2020 (Nostr NIP-01).
How is this different from ensemble methods?¶
Ensembles run N model calls and average the results into a single confident answer. Tractum runs N specialist agents, surfaces disagreement explicitly, and signs the consensus with the trace intact — who said what, where they diverged, how the strategy resolved it.
The receipt includes per-agent responses and the disagreement score, not just the final answer. See the ensemble comparison on the landing page.
How is this different from LangGraph / CrewAI / Goose?¶
Those frameworks coordinate agents — they don't produce signed agreement. Tractum is the consensus primitive those frameworks don't ship. Use them for orchestration; use Tractum when the orchestration needs to land on a verifiable, signed answer.
Adapters exist for LangGraph, CrewAI, MAF, and Nostr transport. Drop Tractum in as one node / tool / Flow.
What does it cost?¶
Rough numbers at Haiku/Sonnet pricing:
| Path | Setup | Cost per consensus |
|---|---|---|
| Cheap (WeightedVote · Haiku · 256 tok) | 3 agents, 1 round | ~$0.01 |
| Debate (Haiku · 1k tok) | 3 agents, 2 rounds | ~$0.30 |
| Premium (Debate · Sonnet · 2k tok) | 3 agents, 2 rounds | ~$1.50 |
100 questions/day at the debate path ≈ $90/month. Use it for what's worth it; skip it when single-model confidence is enough. See Tradeoffs on the landing.
How long does a consensus call take?¶
~4× slower than a single Claude call. The benchmark: 29.8s for a full Tractum Debate vs 7.5s for single Claude (same model, same hard question). The 4× comes from the 2-round Debate path; WeightedVote (1 round, N agents in parallel) is essentially the same latency as a single call.
For interactive UX where 30 s is too long, use WeightedVote instead of Debate.
What happens if Tractum dies? Am I locked in?¶
No. Receipts are self-contained Ed25519 signatures over a canonical JSON payload. Verify them with nak, OpenSSL, or a 20-line Python script — no Tractum runtime required.
The SDK is MIT-licensed and will stay MIT. No BSL rug, no SSPL switch, no Open Core carve-out. If a hosted tier ships later, the SDK stays MIT.
What's the security story?¶
- 17 adversarial tests across 7 attack classes (signature/payload/pubkey/hash tampering; replay; task substitution; single-key compromise; swarm poisoning; canonical-JSON equivocation; Nostr transport tampering).
- No third-party audit yet. The threat model is published at THREAT-MODEL.md — read it before relying on this for compliance.
- Honest "defends / doesn't defend" split: Tractum proves receipt integrity, not agent identity (three "specialists" could be one process in three hats) or agent correctness (confident lies are recorded as confidently as confident truths).
What doesn't Tractum defend against?¶
- Agent identity. Three "specialists" could be one process in three hats.
- Agent correctness. Confident wrong answers are recorded as confidently as confident right ones.
- Signing-key rotation / revocation / forward secrecy. Not implemented yet.
- Post-quantum attacks on default Ed25519/BIP-340. Classical cryptography. The ML-DSA-65 backend ships (
pip install tractum[pq]) for forward-looking attestations. - Nostr relay delivery / ordering. Relays can drop or reorder events. The signature defends integrity, not delivery.
Is there a hosted/managed version?¶
Not yet. Multi-tenancy, billing, Stripe, template marketplace are explicitly deferred until paying demand appears (the "Curtis lesson" — don't build before there's demand). See DEPRECATION.md for the rationale.
A single-tenant runner (FastAPI + SQLite + dashboard, docker compose up) is on the roadmap.
Is it production-ready?¶
v0.3.x — pre-1.0. Public API may shift. 1.0 ships when the signature surface is locked — at that point, canonical payload, receipt shape, and verifier interface freeze.
Currently 164 tests passing, 97% on critical-path modules (signing, transports, transparency, consensus). The MCP server and tools layer are at 0% coverage.
How do I cite this?¶
Tractum — Verifiable multi-agent consensus for Claude.
Jordan Newell, 2026-. https://github.com/JordanNewell/tractum
Or just link to the repo or the docs.
