19 production-grade modules forming the complete economic infrastructure for autonomous AI agents. Every module is formally specified, cryptographically secured, and battle-tested.
Phanes ID
Built on the identity layer for software agents and robots: an issuer-sealed document, a binary Mark for devices, a printed token the browser verifies, a signed record of activity, insurance and certification, and a registry that keeps every request on the record.
An SD-JWT bound to the agent's own key, with document codes IA for software agents and IR for hardware-bound robots. Disclosable claims travel as salted digests; the holder reveals only what a venue needs.
Issued against attestations from model providers, KYB providers and manufacturers; revoked or suspended through issuer-signed status lists that verifiers cache.
A 160-byte binary credential recognised in about 6 ns and verified in about 52 µs by the Rust core; the hybrid form adds a FIPS 204 ML-DSA-87 signature and verifies in about 311 µs.
Rust crate · C ABI (include/phanes_core.h) · JSON-stdio binary · Python binding
The QR on every ID is https://phanes.app/id#<token>. The page reads the token from the fragment, verifies the seal against the issuer key on its own origin, checks the status lists and shows every check. The token never reaches a server.
Served under a Content-Security-Policy that pins the script by hash; any registry serves the same page for its own tokens.
Who, what, when, where, how and why for every action, signed by the agent in a two-phase append and committed to an RFC 6962 Merkle tree with operator-signed heads. Venues witness events inside their visa's scope.
Inclusion and consistency proofs, signed transcripts, and claims that name a proven event.
Four packages (starter, professional, institutional, robotics) with per-claim and aggregate limits, deterministic premiums from the ID and its history, and claims adjudicated from activity-log proofs.
Four certification tiers (Registered, Verified, Certified, Institutional) issued as signed credentials that echo every check they used; the badge travels in HTTP headers and MCP metadata.
Hybrid Ed25519 + ML-DSA-87 seals on every issuer artefact, a hash-chained Merkle audit trail over every request served, per-caller rate limits and security headers, non-custodial signed-request authentication.
Robots apply with a device profile and a manufacturer attestation over measured firmware; the Mark carries the hardware-bound flag.
Economic Layer
The surface where agents become economic actors: verifiable identity, binding contracts, fair dispute resolution, adaptive risk controls, and programmable assets.
DID-based agent identity anchored to Ed25519 keypairs. Selective disclosure lets agents reveal only the attributes required for a given interaction—nothing more.
Delegation chains encode hierarchical authority so supervised agents can act within defined scopes while preserving auditability.
Multi-signature binding agreements with templated creation through ContractFactory: service_agreement, data_purchase, compute_lease.
Escrow supports milestone-based release, automatic penalty enforcement on breach, and deterministic state transitions across the contract lifecycle.
PROPOSED → ACTIVE → COMPLETED / BREACHED / CANCELLED
Three-tier resolution: auto-resolve for clear-cut cases, panel arbitration for structured review, and full tribunal for high-stakes outcomes.
VRF-based random arbitrator selection reduces collusion and predictable forum shopping. Voting is confidence-weighted; evidence carries cryptographic timestamps for integrity and ordering.
Behavioral profiling tracks transaction and interaction patterns over time. Circuit breakers halt suspicious bursts before they propagate system-wide.
Counterparty scoring leverages historical performance; insurance pools provide loss coverage for qualified flows, aligning incentives across the network.
Programmable tokens support configurable decay curves, staking for governance participation, and flexible accrual schedules aligned to real-world settlement cadences.
Token-gated access control ties capabilities to holdings—enforcing least-privilege economics at the protocol boundary.
Intelligence Layer
Real-time anomaly detection, network-level trust inference, and bilateral state channels for micro-transactions—without sacrificing cryptographic guarantees.
Isolation Forest scores live transaction streams for anomalies. Markov chain behavioral models predict agent state transitions from observed sequences.
Entropy drift detection surfaces distribution shifts early—before they become systemic incidents.
PageRank-style trust propagation diffuses reputation across agent graphs. Community detection highlights coordinated clusters for collusion analysis.
Cascade simulation forecasts failure propagation; Sybil detection exploits structural graph signals—not just local heuristics.
Bilateral off-chain payment channels enable high-frequency micro-transactions with minimal on-chain footprint.
Cooperative close settles amicably under normal operation. Force close falls back to on-chain resolution with challenge periods to preserve liveness and safety.
Cryptographic Layer
Secret-shared signing and succinct proofs—implemented for performance, audited for correctness, and bridged into the broader Python stack where needed.
Shamir Secret Sharing over GF(L) with Lagrange interpolation reconstructs secrets only when the threshold is met.
t-of-n threshold signatures evaluate polynomials on shares to produce valid quorum signatures without centralizing key material.
Time-lock puzzles enable delayed reveal flows for escrow, arbitration outcomes, and coordinated upgrades.
Built on the Ristretto255 elliptic curve group with a Merlin transcript-based Fiat-Shamir transform for secure challenge generation.
Delivers zero-knowledge range proofs with a Rust core and Python FFI bridge for ergonomic integration.
Test status: 21 passing tests in the proof subsystem.
Consensus & Settlement
Classical PBFT for replicated state, paired with fiat and stablecoin settlement paths that mirror how production businesses actually move value.
PBFT (Castro–Liskov, 1999) with Ed25519-signed quorum certificates and a full view-change protocol for leader failure.
Tolerates f Byzantine faults among 3f+1 replicas. Cluster sizes from 4–13 nodes are supported for practical deployments.
PaymentIntent escrow, authorize-then-capture: reserve, verify work, capture on success.
Refunds on dispute; webhooks sync ledger with Stripe events.
ERC-20 escrow on Ethereum, Base, Arbitrum, Polygon.
build_approve_tx → build_lock_tx → build_release_tx
Persistence & Audit
ACID storage for operational data paired with an append-only ledger that makes every consequential action auditable—and provable to third parties.
SQLite WAL mode enables concurrent readers while writers commit safely—ideal for embedded and edge deployments without sacrificing throughput.
Full ACID transactions, versioned schema migrations, and robust crash recovery keep operator data consistent across upgrades and failures.
Append-only hash chain: every entry references the previous hash, making tampering detectable instantly.
Merkle tree proofs unlock selective verification—disclose exactly what a counterparty must see, not the entire journal.
A complete audit trail ties protocol events to settlement, disputes, and governance actions.
Integration Layer
Meet agents where they already run: native tool calling for frontier models, OpenAPI-backed services for backends, and a TypeScript SDK for product teams.
35 tools for Claude/GPT-native integration via the Model Context Protocol: 11 for the protocol, 14 for the Phanes ID stack. JSON-RPC over stdio keeps deployment simple and secure.
aeos_create_agent · aeos_create_contract · phanes_id_apply · phanes_mark_verify · …
FastAPI service exposing 87 endpoints with interactive OpenAPI documentation at /docs: 17 protocol routes, 47 non-custodial Phanes ID routes under /v1, 4 public verification routes.
Full CRUD coverage for agents, contracts, and ledger projections; signed-request authentication for everything an ID holder does.
Complete client library spanning crypto primitives, identity lifecycle, and contract authoring helpers.
Typed HTTP client with structured error handling maps protocol failures to actionable exceptions in application code.
With a Phanes ID
One application, keys the agent generates itself, and every module keyed by the same Phanes ID: identity, service-credit settlement, contracts with escrow behind the risk gate, PBFT anchors, disputes with signed evidence and VRF arbitration, an operator-signed ledger, insurance and certification, and a signed inspection report anyone can verify.
Identity
A protocol DID derived from the holder key, a registry membership proof, capabilities and bounds enforced from the sealed document, standing from the status lists.
Contracts and settlement
Holder-signed terms, escrow from service credits at activation, proof digests at fulfilment, release after the anomaly check.
Risk
Assessment at proposal, approval at activation, ML at release, graph metrics and a deterministic reputation.
Consensus
Every operation committed by the replicas; anchors with the exact signed bytes, verifiable in the browser.
Disputes
Two-phase signed evidence, automatic refunds on clear breaches, VRF arbitrators voting with their own keys.
Insurance and certification
Packages with deterministic premiums, claims adjudicated from activity-log inclusion proofs, four tiers that echo every check.
Inspection
A registry-signed, ML-DSA-87-sealed, redacted report of everything a counterparty may know, verified by the CLI, MCP and phanes.app/inspect.
phanes apply ... && phanes credits ... && phanes propose ... \
&& phanes sign-contract ... && phanes fulfill ... && phanes inspect PAYER --verify
# 19 REST routes under /v1, 14 CLI commands, 10 MCP tools; docs.phanes.app/docs/protocol/agent-stackRecorded evidence. Recorded run 20260915T060329Z-evidence: 80 pseudo-terminal sessions at exit 0, a 112-agent campaign with 134 contracts, 20 arbitrated disputes and 83 of 83 inspection reports verified by a third party; every figure is read from the run's files.
Wire the full stack into your agents—identity, contracts, proofs, settlement, and observability—with docs, SDKs, and examples built for production.