Skip to content

MCP × Agent Interaction Architecture

Status: Architecture whitepaper Full version: Chinese whitepaper

This document summarizes Nimi's approach to bidirectional Agent interaction under the Model Context Protocol (MCP).

Core Conclusions

  1. MCP serves as the interoperability protocol layer in Nimi, unifying how Agents connect to tools and capabilities.
  2. Nimi's differentiation is platform-level governance: authorization, scope convergence, fail-close semantics, and audit evidence chains are enforced as strong constraints.
  3. For App/Mod developers using the Nimi SDK or nimi-hook, governance sinks into the platform by default — low integration cost, high observability.

Interaction Model

Inbound (External → Nimi)

External Agents invoke Runtime-exposed capabilities via MCP following a fixed execution protocol:

discover → dry-run → verify → commit → audit

Each stage enforces authorization, scope validation, and audit trace. High-risk write operations require the full chain; skipping verify before commit triggers fail-close rejection.

Outbound (Nimi → External)

The Runtime acts as an MCP Client calling external MCP Servers (search, prompt sources, enterprise tools). Connection lifecycle includes config-driven hot reload, exponential backoff on failure, and circuit-breaking after 3 consecutive failures.

Governance Stack

  • Auth/Grant: Principal model (Human / NimiAgent / ExternalAgent / Device / Service), session TTL bounds, scope prefix validation, delegation with depth limits.
  • Audit: Minimum 6 fields on every audit path (trace_id, app_id, domain, operation, reason_code, timestamp). Sensitive fields are redacted at the audit write layer.
  • Fail-close: Insufficient authorization, missing verification, or uncertain persistence all result in rejection.

Developer Experience

App/Mod developers declare Action schemas, capabilities, and handlers. Authorization, audit, and error projection are managed by the platform (Runtime / SDK / nimi-hook). The primary new cost is action modeling quality — schema, risk classification, and compensation semantics.

Spec Rule Anchors

  • Platform: P-ALMI-*, P-ARCH-*, P-PROTO-*
  • Runtime: K-AUTHSVC-*, K-GRANT-*, K-AUDIT-*, K-CONN-*, K-ERR-*
  • Desktop: D-SEC-*, D-HOOK-*, D-MOD-*, D-IPC-*

Released under Apache-2.0 (runtime, sdk, proto) / MIT (desktop, nimi-mods, web) / CC-BY-4.0 (docs)