Skip to content

ADR-0036 — The LangChain Ecosystem Reproduces the Cloud Platform MCP (Self-Hosted)

Status

Accepted — Research + adoption plan (September 2026) Repository: https://github.com/FinnMacCumail/ollamaDeepAgents (docs/development/2026-08-10_langchain-ecosystem-vs-netbox-cloud-platform.md)

Context

NetBox Labs' Cloud-only Platform MCP Server advertises an "agent-native" bundle: ~100 tools, sandboxed Code Mode, dynamic model discovery, tier-enforced tool registration, an Agent Skills knowledge layer, and structured self-correcting errors. The question 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? A five-agent survey mapped each Cloud capability to its closest open-source, self-hostable equivalent.

Decision

Reproduce the Cloud Platform MCP's query-processing value from open-source, self-hostable pieces; skip every hosted/paid convenience under the privacy mandate.

Capability-by-capability verdict:

Cloud capability Self-hostable reproduction Verdict
Code Mode (sandboxed executor) langchain-quickjs PTC Skip — reconfirms ADR-0032; Anthropic's own τ²-bench shows sequential single-call workloads gain nothing (~8% more cost). The GraphQL tool is the right lever for round-trips.
The round-trip problem Code Mode targets Read-only GraphQL tool (ADR-0034) Already shipped — server-side join
~100 tools without context bloat LLMToolSelectorMiddleware · DeepAgents subagent partitioning · langgraph-bigtool Available, native — deferred until 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, #33808)
Structured self-correcting errors RubricMiddleware + structured errors emitted by our MCP server High value — the OSS equivalent
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 Aug-2026 LangChain "SRE agent" template) The blueprint for the next feature

Skip (hosted/paid, fail the privacy mandate): Managed Deep Agents (US-cloud only), LLM Gateway (no local provider), Context Hub (use own git repo), hosted LangSmith. Self-hosted eval/observability parity comes from the MIT openevals / agentevals packages + OpenTelemetry export.

Consequences

Positive

  • The Cloud product's genuinely novel agent-facing ideas — portable Skills and Code Mode — are the easiest to reproduce (Skills already in use; Code Mode is the already-investigated PTC pattern). The hard part is tool breadth, which the read-only workload does not yet need.
  • Produced a concrete, prioritized adoption plan: (1) the deepagents 0.7.5 upgrade ([ADR-0035]), (2) model-handoff routing via subagents, (3) RubricMiddleware, (4) strengthen NL→query.
  • Everything on the plan is OSS and on-prem — consistent with the strict data-privacy mandate.

Negative / limitations

  • The Cloud product's managed convenience (hosting, tiered RBAC sessions) is genuinely not reproducible self-hosted — but it is orthogonal to query-processing quality.
  • LangChain cannot add/remove tools after agent creation (#33808), so "dynamic discovery" must be done server-side at startup, not continuously as the Cloud product does it.

Relationship to prior ADRs

Reconfirms ADR-0032 (Code Mode / PTC deferral) from the ecosystem-wide vantage; triggered ADR-0035 (the 0.7.5 upgrade); and frames ADR-0034 (GraphQL) as the correct lever for the round-trip problem Code Mode claims to solve.

References