Building the HTTP for the Age of Generative AI
Current version: v1.9.0
Reference implementation: iicp.network
Status: Active development — Phase 5 (Cooperative Inference Protocol)
IICP is an open protocol that lets AI agents discover each other, negotiate capabilities, and route tasks across a distributed network — without any central broker owning the compute or controlling the data.
Agent A ──CALL──▶ IICP Node B ──▶ LLM Backend
▲
│ discovery
iicp.network
(directory only —
no payload passes through)
The directory (iicp.network) is bootstrap and discovery only and does not receive task
payloads. Tasks route to the selected execution node, whose operator can read the work it executes.
Current IICP-CX clients encrypt requests across the network and relays when a provider advertises
cx_public_key; this is transport confidentiality, not executor-blind inference or anonymity.
{
"task_id": "550e8400-e29b-41d4-a716-446655440001",
"intent": "urn:iicp:intent:llm:chat:v1",
"payload": { "messages": [{ "role": "user", "content": "Summarise this doc." }] },
"constraints": { "timeout_ms": 5000, "qos": "interactive" }
}An intent URN expresses what you want, not which model or endpoint to call. The network finds the best available node that can serve that intent, routes the task, and returns a structured response.
| Document | What it covers | Normative level |
|---|---|---|
| iicp-core.md | Wire format, 14 message types, required fields, error codes, security minimums | MUST |
| iicp-semantics.md | Intent routing, QoS tiers, node scoring, retry policy, circuit breaker | SHOULD / MAY |
| iicp-extensions.md | Billing, reputation, MCP binding, Cooperative Inference, post-quantum | MAY / future |
| Document | What it covers |
|---|---|
| iicp-dir.md | IICP-DIR: directory registration, heartbeat, discovery, peer exchange |
| iicp-mcp-binding.md | MCP ↔ IICP translation rules |
| node-capability-format.md | Capability object schema: intents, models, limits, availability |
| iicp-billing-extension.md | Credits, billing fields, receipt protocol |
| iicp-cbor-wire.md | Optional CBOR wire encoding (Phase 3, application/iicp+cbor) |
| Document | What it covers |
|---|---|
| IICP-core-phase1-profile.md | Phase 1 field subset — the minimum viable implementation |
| conformance-test-suite.md | 200+ machine-verifiable test IDs (DIR-, PROXY-, SEC-, CIP-, DIR-FED-*) mapped to REACH probes |
| validation-methodology.md | How to validate implementations; performance claim disclosure |
| iicp-v1.5-overview.md | What changed in v1.5; migration guide from v1.4.2 |
| Path | Contents |
|---|---|
| schemas/task.json | JSON Schema 2020-12 for IicpTask |
| schemas/nodelist.json | JSON Schema 2020-12 for NodeListResponse |
| registry/intents.json | Official intent URN registry |
| spec/intent-risk-taxonomy.json | Shared prohibited/high-risk/transparency/minimal intent classification fixture |
| spec/mcp-tool-risk-taxonomy.json | Shared MCP tool-risk and default-gating fixture |
| Path | Contents |
|---|---|
| spec/archived/IICP_draft_1.4.2.txt | Original monolithic Internet-Draft (archived) |
Read IICP-core-phase1-profile.md for the minimum field subset. A Phase 1 node must:
- Register —
POST /v1/registerwith{endpoint, region, capabilities[], limits} - Heartbeat —
POST /v1/heartbeatevery 30 s with{load, active_jobs} - Accept tasks —
POST /v1/task— validate UUID-v4task_id, intent URN,timeout_msbounds - Discover peers —
GET /v1/discover?intent=urn:iicp:intent:llm:chat:v1 - Return structured errors — never raw exceptions; always
{"error": {"code": "IICP-Exxx", ...}}
| Level | Documents to satisfy | Typical implementer |
|---|---|---|
| Core | iicp-core.md MUSTs only | Minimal node, embedded device |
| Phase 1 | Core + IICP-core-phase1-profile.md | Reference implementation |
| Phase 2 | Phase 1 + iicp-dir.md §3.6 (peers) | Mesh node |
| Phase 3+ | Phase 2 + billing/reputation extensions | Full CIP node |
Three official client SDKs (current release: v0.7.87) implement
both sides of the protocol — the consumer (discovery, routing, retry, fallback, CIP
consumer) and the provider (iicp-node runtime with backend auto-detection, NAT
escalation, relay worker/server modes, and a built-in MCP gateway). All are open-source
and published:
| Language | Install | Package registry | Source repository |
|---|---|---|---|
| Python | pip install iicp-client |
PyPI: iicp-client | github.com/RobLe3/iicp-client-python |
| TypeScript | npm install @iicp/client |
npm: @iicp/client | github.com/RobLe3/iicp-client-typescript |
| Rust | cargo add iicp-client |
crates.io: iicp-client | github.com/RobLe3/iicp-client-rust |
No install at all? iicp.network/browser-node runs a real model in your browser (WebGPU) and queries the live mesh as an IICP consumer — with a connection console that shows every discover/dispatch wire step.
The SDKs are conformant reference clients — a good starting point for understanding the wire format in practice. Bug reports and PRs are welcome on each repository.
A provider node should become reachable without router surgery. The SDKs escalate
automatically: direct endpoint → UPnP pinhole → IPv6 GUA → relay
auto-election from the directory (outbound bind; transport_method=turn_relay) →
Quick Tunnel (zero-account cloudflared; transport_method=external_tunnel) —
each rung tried only when the previous fails, each surfaced honestly in the node's
exposure_mode so discovery and scoring can see how a node is reached.
The protocol's normative choices are backed by simulation and analysis. The full research
record — credit economy & rate calibration, reputation/tier modelling, adversarial robustness
(FRAME8, REP, MESH), routing/multi-path selection, cryptographic trustworthiness, portable
operator identity, gamification anti-gaming, NAT traversal, and more — lives under
research/, indexed in research/RESEARCH.md.
These notes are published so the spec's decisions are externally verifiable: each major parameter (tier weights, credit schedule, decay floors, EMA α, etc.) traces back to a documented simulation or analysis. Found a flaw or have a better method? The research is meant to be challenged — open an issue.
The live reference implementation also keeps a public research summary at iicp.network/research. Treat live-network evidence, controlled validation, simulation and future research as different confidence levels; do not cite simulations as production measurements.
Intent URNs identify what is being requested, independent of model or backend:
urn:iicp:intent:<domain>:<action>:v<version>
Examples from registry/intents.json:
| URN | Purpose |
|---|---|
urn:iicp:intent:llm:chat:v1 |
Conversational LLM completion |
urn:iicp:intent:llm:embedding:v1 |
Text embedding / vector |
urn:iicp:intent:llm:summarise:v1 |
Document summarisation |
urn:iicp:intent:vision:describe:v1 |
Image-to-text description |
urn:iicp:intent:audio:transcribe:v1 |
Speech-to-text |
To propose a new intent, open an issue with the URN, domain justification, and example payload.
The directory scores nodes server-side at query time. Clients receive a pre-sorted list and may only filter (remove available=false nodes or those with open circuit breakers).
Phase 3 scoring formula (currently deployed):
score = 0.35 × availability_factor
+ 0.28 × (1 − normalized_load)
+ 0.18 × capacity_ratio
+ 0.09 × region_match
+ 0.10 × reputation_score
See iicp-semantics.md for full term definitions.
All IICP implementations MUST:
- Use TLS 1.3 minimum for all inter-node communication
- Validate
task_idas UUID v4 - Validate
intentagainsturn:iicp:intent:[a-z0-9:]+:v[1-9][0-9]* - Validate
timeout_msin range [100, 300 000] - Never log or expose
payloadcontent - Return structured errors only — no stack traces, no filesystem paths
See conformance-test-suite.md SEC-* test IDs for machine-verifiable checks.
Cooperative Inference and routing hardening (active)
The iicp.network directory is live and the client SDKs are published at v0.7.87 (PyPI / npm / crates.io). Each includes the iicp-node provider runtime, so a participant can use the mesh first and provide capacity later. Live node availability, installed-version adoption and encryption evidence change over time; consult the live stats page before treating any network condition as current.
The mesh is usable for its current capabilities, while relay hardening, broader privacy evidence and public federation remain separate maturity gates. Remote execution still means the selected provider can read the task it executes.
| Feature area | Status | Notes |
|---|---|---|
| Core protocol — register / discover / route | ✅ Live | Current evidence is published on the live stats page |
| CIP coordinator (multi-node dispatch) | ✅ Implemented | Credit receipts, response integrity verification |
| Reputation scoring | ✅ Ratified | Tier structure (§5.1.1) + bootstrap floor (§5.1.2) ratified 2026-05-24 — normative |
| Published SDKs (Python / TypeScript / Rust) | ✅ Published v0.7.87 | Shared protocol baseline; see Client SDKs |
Node runtime (iicp-node) |
✅ Published | Ships inside every SDK (pip install iicp-client → iicp-node serve) |
| Relay transport for unreachable workers | ✅ Shipped (v0.7.56) | HTTP long-poll worker transport — browsers and CGNAT operators bind outbound to a relay-capable node; consumers route through path-scoped relay endpoints with zero client changes |
| Browser node (WebGPU, zero install) | ✅ Live | iicp.network/browser-node — runs a real model in the browser via WebLLM, queries the live mesh as an IICP consumer (with a wire-level connection console), and can serve into the mesh via a relay. First directory-listed browser node verified end-to-end on 2026-06-12 |
| Browser-consumable nodes (CORS) | ✅ Shipped (v0.7.56) | Every node endpoint answers CORS preflights — any https-exposed node can serve web-page consumers directly |
| Automatic NAT escalation incl. Quick Tunnel | ✅ Shipped (all 3 flavours) | Ladder: direct → UPnP → IPv6 → relay auto-election → zero-account Cloudflare Quick Tunnel fallback. Direct paths stay preferred; Quick Tunnel remains a low-friction bootstrap/fallback rather than a production availability promise |
| Signed event log + compliance attestation | ✅ Live | Every registration/heartbeat/eviction in a cryptographically signed log (federation bootstrap source); signed compliance attestation endpoint |
| Federation (Phase 6 groundwork) | 🟢 FED-READY-1 proven | Rust replica directory bootstraps from the PHP seed via snapshot + signed event tail |
| Operator identity (Ed25519 delegation) | 🟢 Phase A live | ADR-045 — operators sign a delegation binding their Ed25519 key to each node; the directory verifies + resolves a public operator_display_name in discovery. operator_pubkey is directory-private, never served. |
| Founder recognition | 🟢 Live | Time-gated founder ordinals (iicp-recognition §5.4) — #1 reserved for the maintainer, #2..N earned by genuine served nodes; dedicated non-federated signed chain |
Estimated progress toward closed beta: ~80%
The mesh works end-to-end, the SDKs are publicly installable with full three-language parity, and a browser tab can both consume the mesh and (via relay) serve into it. The remaining gaps before closed beta are:
- A standing public relay (the transport is built and verified; one reachable host activates browser/CGNAT serving for everyone)
- A portable operator identity wallet (so node identities survive machine changes)
- Security and authentication hardening to production standard
- Complete live SDK/key adoption and verified privacy receipts before strict fail-closed privacy wording
Follow this repo or iicp.network for announcements.
| Version | Date | Notes |
|---|---|---|
| v1.9.0 | 2026-05-30 | Security-hardening normative content: per-heartbeat reputation delta cap (§11.2), audit-report griefing cap (§11.5). Directory drift closeout: AUDIT_REPORT endpoint, Public Stats schema, free-credit rules, NODELIST health_label/exposure_mode/public_key + transport fields; credit-endpoint / SCORE_UPDATE-snapshot / tier-enum reconciliations |
| v1.8.0 | 2026-05-25 | S.13 ephemeral-by-design federation (ADR-033): HEARTBEAT/SCORE_UPDATE/REPUTATION_UPDATE removed from federated event log; snapshot+event-tail bootstrap (GET /v1/snapshot); replica registration handshake |
| v1.7.0 | 2026-05-24 | §5.1.1 tier structure + §5.1.2 bootstrap floor ratified; all 13 Phase-5 research tracks closed; mesh-health compound metric live |
| v1.6.0 | 2026-05-23 | CIP receipt response integrity (TC-9c response_hash); framing spec stubs (CBOR/QUIC); 12 additional sub-protocol docs |
| v1.5.0-draft | 2026-05-15 | Spec split (core/semantics/extensions); 7 sub-protocol docs; 40 conformance test IDs; CBOR wire format |
| v1.4.2 | 2024 | Original monolithic Internet-Draft |
v1.5+ is wire-compatible with v1.4.2 — no message opcodes, field names, or error codes changed.
The reference implementation (RobLe3/iicp.network) is currently private — it will be opened once the system reaches public beta readiness. The architecture is:
| Component | Language | Purpose |
|---|---|---|
directory/ |
PHP 8.3 + Laravel | Control plane — registration, discovery, scoring (interim, shared-hosting) |
iicp-directory-rs/ |
Rust 2021 + axum | Control plane — permanent replacement for the PHP directory (auditable scoring compiled into one binary) |
adapter/ |
Python 3.11 + FastAPI | Execution plane — task acceptance, backend dispatch |
proxy/ |
Python 3.11 | Client plane — discovery, routing, retry, fallback |
iicp-node/ |
Rust 2021 + tokio | High-performance node runtime |
The protocol specification in this repository is the authoritative source for building interoperable implementations. Third-party implementations that conform to the spec (see conformance test suite) are fully compatible with the live network.
- Bug reports / clarifications: Open an issue
- New intent URNs: Open an issue with URN, domain, and example payload
- Protocol proposals: Open an issue tagged
protocol-change - Conformance tests: PRs to
spec/v1.9/conformance-test-suite.mdwelcome
All normative language follows RFC 2119 / BCP 14.
| File | Purpose |
|---|---|
| tools/protocol_integrity_analysis.py | Analyses a spec file for internal consistency |
| tools/quick_validation.py | Quick syntax + field validation against v1.4.2 |
