Skip to content

Architecture Review: Consolidated Stack

Status: Revised (2026-02-02) Date: 2026-02-02

Executive Summary

The consolidated stack is structurally sound and aligns with the goals of protocol-agnostic tooling, progressive discovery, and sandboxed execution. The largest risks are runtime backend completeness and protocol adapter coverage. Auth consolidation is complete and now lives in toolops/auth.

What Is Solid

  • Layer separation is clean: toolfoundation → tooldiscovery → toolexec → toolcompose → toolops → toolprotocol → metatools-mcp.
  • Canonical model embeds MCP SDK for spec fidelity.
  • Progressive discovery has a clear pipeline and interfaces.
  • Execution/runtimes are isolated behind stable contracts.

Gaps to Close (Priority Order)

  1. Protocol adapters beyond MCP/OpenAI/Anthropic (A2A + Google Gemini).
  2. Runtime parity for Kubernetes, gVisor, Kata, Firecracker, and remote backends.
  3. Auth consolidation: complete (single source in toolops/auth).
  4. Spec alignment: MCP 2025-11-25 feature coverage and testing.

Decisions (Reaffirmed)

  • Keep CanonicalTool as the hub-and-spoke adapter format.
  • Keep protocol primitives in toolprotocol; do not fork per protocol.
  • Treat metatools-mcp as a reference server, not the source of truth.

References

  • ai-tools-stack/docs/roadmap.md
  • ai-tools-stack/docs/architecture/stack-map.md
  • ai-tools-stack/docs/architecture/protocol-crosswalk.md