Phase 5: Reproducing the Cloud "Platform MCP Server" on the Open-Source LangChain Stack¶

Research Overview¶

Research Question: NetBox Labs sells a Cloud-only Platform MCP Server that makes infrastructure "agent-native" — ~100 tools, a sandboxed Code Mode, dynamic model discovery, tier-enforced tool registration, an Agent Skills knowledge layer, and structured self-correcting errors. For a self-hosted, read-only, privacy-mandated build: how much of that is reproducible on the current (Aug 2026) LangChain / LangGraph / DeepAgents / LangSmith ecosystem — and what should actually be adopted?

Method: five parallel deep-research agents, each mapping one dimension (Code Mode, large tool sets, routing/handoff, current DeepAgents/blog, NL→structured-query) against the Cloud product, all primary-source-cited.

Finding (short version): the query-processing value is largely reproducible from open-source, self-hostable pieces — and the Cloud product's two genuinely novel ideas (portable Skills, Code Mode) are the easiest to reproduce. What is not reproducible is the managed convenience, which is orthogonal to answer quality. Every hosted/paid piece is skipped under the privacy mandate.

Repository: https://github.com/FinnMacCumail/ollamaDeepAgents (docs/development/2026-08-10_langchain-ecosystem-vs-netbox-cloud-platform.md)

Capability-by-capability map¶

Cloud capability	Closest self-hostable reproduction	Verdict
Code Mode (sandboxed executor, "75% fewer round-trips")	`langchain-quickjs` PTC middleware	Skip — see below
The round-trip problem Code Mode targets	Read-only GraphQL tool (GraphQL Read Path)	Already shipped
~100 tools without context bloat	`LLMToolSelectorMiddleware` · DeepAgents subagent partitioning · `langgraph-bigtool`	Native; deferred until the tool surface grows
Dynamic model discovery	Generate tools from NetBox OpenAPI/GraphQL at MCP-server startup	Server-side (LangChain can't add tools mid-run)
Tier-enforced tool registration	Filter the tool list per role at agent-build time	Static-per-session
Structured self-correcting errors	`RubricMiddleware` + structured errors from our own MCP server	The OSS equivalent — high value
Agent Skills knowledge layer	DeepAgents Skills (open `agentskills.io` spec)	Already in use
NL → correct structured query	Curated GraphQL operations + schema-RAG + tool-arg validation	Highest correctness leverage
Model routing / handoff	DeepAgents subagent `task()` (the LangChain "SRE agent" template)	The blueprint for the next feature

The headline: Code Mode is the wrong lever for this workload¶

The Cloud product's "75% fewer tool round-trips" comes from its ~100-tool catalog + batch tasks — neither of which a read-only, single-source, dependency-chained NetBox workload has. This independently reconfirms the earlier QuickJS/PTC deferral (ADR-0032): Anthropic's own τ²-bench shows that sequential single-call workflows gain nothing from programmatic tool calling and cost ~8% more — matching this project's own spike (+13.7% latency, model never chose the code tool). The right lever for the round-trip problem was the read-only GraphQL tool (a server-side join), not a sandbox.

What's worth adopting (all open-source, all on-prem)¶

The research produced a prioritized plan, sequenced by what unblocks what:

DeepAgents 0.7.5 upgrade — tracks the subagent + RubricMiddleware APIs the rest needs. Done: DeepAgents 0.7.5 Upgrade / ADR-0035.
Model-handoff routing via DeepAgents subagents — the LangChain "autonomous SRE agent" (Aug 2026) is almost the exact architecture: tiered models by complexity, tool-scoped subagents. A fast local model fields simple queries; a heavier model + GraphQL subagent handles multi-hop.
RubricMiddleware — structured self-correction gated on project criteria (the OSS parallel to the Cloud's "self-correcting errors").
Strengthen NL→query — curated cross-domain GraphQL operations + tighter tool-arg schemas.

What's skipped (hosted/paid — fails the privacy mandate)¶

Managed Deep Agents (US-cloud only), the LLM Gateway (no local-model provider), Context Hub (a git repo serves the same purpose), and hosted LangSmith. Self-hosted eval/observability parity comes from the MIT openevals / agentevals packages plus OpenTelemetry export — the project already runs its model-matrix eval and reference-grounded correctness judge locally, so this is a natural fit.

Takeaway¶

The "agent-native" bundle a vendor sells as a managed product decomposes, on inspection, into pieces that are mostly open and self-hostable — with the genuinely novel agent-facing ideas (Skills, Code Mode) being the easiest to reproduce and the managed hosting being the only thing you truly can't rebuild. For a privacy-mandated build, that is the whole point: frontier-grade agent capability and data that never leaves the network are not mutually exclusive — they are an engineering plan.