SIPswarm is the first Session Border Controller designed from the ground up with LLM-first management and real-time bid/ask provider negotiation. Swarm architecture. Infinite resilience. No big iron required.
SIPswarm rethinks the Session Border Controller from first principles. No legacy assumptions. No inherited limitations.
The first SIP switch where AI is the primary management interface. Ten AI agents per node handle configuration, security, routing optimization, fraud detection, and infrastructure monitoring. The AI managing the platform uses the same infrastructure it manages. Self-referential by design.
LLM-First ManagementInspired by Starlink satellite mesh and military MANET radios. Many small identical nodes replace big expensive switch pairs. No master, no standby, no controller. Every node is expendable. Add a cloud VM in 5 minutes vs. buy big iron with 3-month lead time. Mix cloud and bare metal in one fabric.
Mesh TopologyRedundancy through quantity, not indestructibility. Lose 1 of 12 nodes and you lose 1/12 capacity -- never a platform-wide outage. Sessions migrate invisibly between nodes via SIP re-INVITE. Percolation theory: highly connected meshes can lose 60-70% of nodes before fragmentation.
Linear DegradationTraditional SBCs deploy as two expensive boxes in active/standby or active/active HA. One fails, the other catches everything -- or doesn't. That model is brittle, expensive, and doesn't scale.
SIPswarm replaces that with a mesh of identical, expendable nodes. Every node runs the same software. Every node handles real traffic. There is no special node. Add capacity by adding nodes. Shrink by removing them. Mix low-cost cloud VMs with on-premise hardware in the same fabric.
Modeled on stock exchange trading mechanics. One inbound request, simultaneous bid to multiple providers, best provider wins. Terms negotiated in real-time during session setup using custom SIP headers. Fully auditable. Every round logged immutably.
One inbound request fans out simultaneously to multiple providers. Each responds with an ask price. Best price wins. MaxBidRnd bounds rounds to control latency. Rate sheet fallback when providers don't support bid/ask.
Bilateral proof of rate agreement. Terminator generates on acceptance, originator echoes in ACK and BYE. Immutable audit trail for every transaction.
The same bid/ask framework applies to AI model sessions, voice/video media, and SMS. One negotiation protocol for all three traffic types.
Every bid log and CDR is SHA-256 hash-chained to its predecessor. Modifying any record breaks every subsequent hash. Cross-node Merkle trees verify fabric-wide integrity. Tamper-proof by math, not policy.
UDP port 5060. Standard SIP. Fourteen custom headers. That is the entire external surface. Admin traffic, media, and data all live on separate internal planes.
UDP :5060 is the only externally-facing port. Admin SIP on :15060 is internal-only with mTLS. Media ports are dynamically allocated and topology-hidden via SDP rewrite.
14 custom SIP headers carry bid price, ask price, negotiation ID, receipt GUID, session type, currency, and round management. Standard stacks ignore them — zero breaking change.
Mode 1 (B2BUA): full SBC in signaling and media path. Mode 2 (302 Redirect): routing advisory only, customer's switch connects directly. Same LCR, same audit.
Rate sheet rate is locked before INVITE is sent. If the carrier ignores X-BF-* headers, the default price already applies. No protocol negotiation. No PDD penalty.
Propose in 200 OK, confirm in ACK echo, carry in BYE. Bilateral proof of negotiated rate. Void path for parallel auction losers (ACK without echo + immediate BYE).
Direct path (bypass SIPswarm), market path (full LCR + bid/ask + CDR + audit), or hybrid. Involvement is earned, not forced. Like NYSE vs OTC — customer chooses.
When a node fails in SIPswarm, traffic redistributes across the remaining mesh. Active sessions migrate invisibly via SIP re-INVITE -- endpoints never know it happened. This is not failover. This is a swarm absorbing a loss.
Percolation theory proves that highly connected meshes can lose 60-70% of nodes before the network fragments. A 12-node swarm losing one node loses exactly 1/12 capacity. No cascade. No outage. No pager going off at 3 AM.
From deployment to traffic. Four steps. No big iron. No long procurement cycles.
Spin up cloud VMs, rack bare metal, or both. Every node runs the same image. No configuration differences. Takes minutes.
Nodes discover each other via SIP REGISTER. The swarm self-organizes into a fully connected mesh. Admin coordination runs on SIP over mTLS.
Ten agents per node activate: configuration, routing optimization, security, fraud detection, infrastructure monitoring, provider health, ETL, and protocol diagnostics.
AI model sessions, voice/video, SMS. Real-time bid/ask negotiation. LCR-optimized routing. Rated CDRs generated for every transaction. Financial-grade audit.
The economics and architecture are fundamentally different.
| Traditional SBC | SIPswarm | |
|---|---|---|
| Architecture | 2 boxes, active/standby | N nodes, active/active mesh |
| Time to start | Months of procurement | Minutes to deploy |
| Scaling | Buy bigger box (months) | Add a node (minutes) |
| Failure mode | Failover (all or nothing) | Linear degradation (lose 1/N) |
| Management | CLI / Web GUI / NOC staff | LLM-first (AI agents) |
| Pricing | Static rate sheets | Real-time bid/ask negotiation |
| AI traffic routing | Not designed for it | Primary traffic type |
| Infrastructure | Dedicated hardware only | Hybrid cloud + bare metal |
| Audit trail | Database logs | SHA-256 hash chain + Merkle verification |
Every node in the fabric runs a full complement of specialized AI agents. A Manager coordinates intra-node operations. A Chief of Staff handles fabric-wide decisions via leaderless quorum consensus. Eight domain agents cover every operational concern.
The AI managing the platform uses the same infrastructure it manages. The LLM agents that configure routing and monitor health are themselves routed through the same bid/ask negotiation engine. Self-referential by design.
"AI improves the rules. The rules run the traffic."
SIPswarm handles three traffic types through one unified platform: AI model sessions, voice/video media, and SMS. SMS rides the same bid/ask negotiation, LCR routing, and CDR pipeline as every other traffic type.
The SMS gateway speaks SMPP (Short Message Peer-to-Peer) — the industry-standard protocol connecting application platforms to carrier message centers (SMSCs). Real protocol. Real delivery receipts. Real carrier integration.
SIPswarm generates rated CDRs in the transaction path itself — not a sidecar, not a log scraper, not a downstream system. The record is born in the same process that handles the call. Then it passes through four layers of integrity protection before it ever leaves the node.
CDRs are created in the same Rust process that handles the SIP transaction. No log scraping. No sidecar. No separate system that might miss a record.
Every CDR and bid log entry carries a SHA-256 hash of the previous record. Alter one record and the entire chain breaks — visibly, immediately, and irreversibly.
Periodic Merkle trees computed per node. Roots exchanged via admin SIP PUBLISH. Any discrepancy between nodes triggers an immediate audit alert.
Bilateral proof of the negotiated rate. The terminating carrier generates a GUID on acceptance; the originator echoes it in ACK and BYE. Dispute evidence built into the protocol.
SIPswarm implements 35+ industry standards and RFCs. No proprietary protocols. No vendor lock-in. Every layer is standards-compliant and auditable.
14 custom SIP headers for real-time price negotiation, defined per RFC 3261 Section 7.3.1 extension mechanism. Fully backwards-compatible with standard SIP infrastructure.
SIPswarm is purpose-built for the AI era. Real-time pricing. Swarm resilience. LLM-first management. No big iron. No legacy baggage.