MCPL is a backward-compatible extension to the Model Context Protocol (MCP) that enables servers to be active participants in the inference lifecycle.
- Push Events — Servers can push events to hosts that may trigger model inference
- Context Hooks — Servers can inject context before inference, with observation and injection granted separately
- Server-Initiated Inference — Servers can request autonomous inference from the host
- Capability Grants — Host-controlled authorization for server behaviors
- Feature Sets — Named behavior bundles derived from the capability grant
- Event Tags — Semantic labels for host-controlled event treatment
- Manifest Changes — Hosts re-fetch and compare announced server changes
- WebSocket Transport — Persistent bidirectional connections with the same capability negotiation as other transports
- Endpoint URIs —
mcpl://locators resolve to secure WebSocket endpoints - Bulk Content References — Bounded references to large payloads, with host-controlled fetching and context inclusion
See SPEC.md for the full protocol specification.
WebSocket transport is adopted separately from the state/branches proposal; see the accepted transport proposal.
Draft specification (v0.5.0-draft). Subject to change. SPEC.md is the authoritative integrated text; RFC files preserve proposal and review history.
| RFC | Status | Integrated specification |
|---|---|---|
| 001: Event Tags | Accepted; incorporated | §16 |
| 002: Capability Grants | Accepted; incorporated | §5.4, §6.7, §14 |
| 003: Manifest Changes | Accepted; incorporated | §17 |
| 004: Endpoint URIs | Accepted; incorporated | §18 |
| 005: Bulk Content References | Draft; incorporated, acceptance pending executable vectors | §19 |
RFC-005's incorporation into the draft does not waive its acceptance criterion:
freeze executable vectors under conformance/ and run them against a strict schema/parser
implementation and a host-treatment implementation.
MIT