NetBox Labs Platform MCP Server

The Platform MCP Server is a hosted Model Context Protocol server (on NetBox Cloud) that puts a NetBox instance behind an MCP tool surface so an agent can read and modify NetBox data. It is distinct from the open-source netbox-community/netbox-mcp-server: the platform server adds Code Mode, per-user auth, dynamic plugin/model detection, and first-party product integrations (Assurance, Discovery).

Your knowledge of this server may be outdated. It is Public Preview; tool inventories, helper signatures, and available namespaces are generated live per deployment and evolve. Always discover the real surface (schema search / tool list) before acting — never assume an object type, endpoint, or helper exists.

Retrieval Sources

Connect & Auth

• Endpoint: https://<instance>.cloud.netboxapp.com/mcp
• Transport: MCP over streamable HTTP (no stdio — it is a networked, hosted server, not a local subprocess).
• Auth: Authorization: Bearer nbt<token> — a NetBox v2 API token (prefix nbt, available on NetBox 4.5+). With per-user auth, your own token's NetBox RBAC applies to every call.
• Customers enable the server on their instance via NetBox Labs support.

Any MCP-over-streamable-HTTP client works (Claude Code/Desktop, Cursor, VS Code, ChatGPT Developer Mode). The docs page has copy-paste client configs.

FIRST: Identify Your Mode

The server exposes tools in one of three modes; what you see in the tool list tells you which mode you're in:

• Code Mode — you see exactly three tools: netboxsearchschema, netboxexecute, rediscovernetbox. Higher tiers default here. This is the primary, most efficient surface.
• Discrete Mode — you see ~34 netbox_* tools (plus plugin/product tools). Each operation is its own tool.
• Both — you see all of the above.

If you see netboxexecute, use Code Mode. If you only see discrete netbox* tools, jump to [Discrete Mode](#discrete-mode).

Code Mode

Instead of dozens of discrete tools (one round-trip each, bloated context), Code Mode exposes three tools whose arguments are Python code. You write a short program that calls injected NetBox helpers, does its own filtering/joining/looping in-process, and returns a single result. This collapses many tool round-trips into one.

The three tools

The golden loop (minimize round trips)

1. Search the schema first — don't guess. Call netboxsearchschema to confirm the right object type, filter field, or special endpoint. Guessing object types/paths is the #1 failure mode.
2. Write ONE netboxexecute program that does the whole job — fetch, filter, join, loop, compute in a single program. Never call netboxexecute once per object when a loop or one list call would do.
3. Always assign result. If you never assign it you get back "Code ran successfully but did not assign to result" — a wasted turn.
4. Project fields with fields=[...] on reads to keep payloads (and the result-size cap) small.
5. Prefer named helpers over raw api() — they carry validation and clearer errors.

# netbox_search_schema — confirm the type/filters first
result = {k: v for k, v in schema['object_types'].items() if v['app'] == 'dcim'}

# netbox_execute — do the whole task in one program
sites = get('dcim.site', filters={'name': 'NYC-DC1'}, fields=['id', 'name'])
site_id = sites['results'][0]['id']
devs = get('dcim.device', filters={'site_id': site_id, 'status': 'active'},
           fields=['id', 'name', 'role', 'primary_ip4'], limit=1000)
result = {'site': 'NYC-DC1', 'count': devs['count'], 'devices': devs['results']}

The sandbox contract (what your code may do)

netboxexecute / netboxsearch_schema run through a validated sandbox. Respect these or your code is rejected:

• NO import / from ... import. This is the single most common failure — agents reach for json, re, ipaddress, math, collections. Everything you need is already in the namespace or in allowed builtins. Never write import.
• NO class definitions, del, async/await/yield.
• NO dunder / introspection (eval, exec, open, getattr, globals, class, subclasses, etc. are blocked).
• Allowed builtins only — the common ones: len, str, int, float, bool, list, dict, set, tuple, range, enumerate, zip, sorted, sum, min, max, any, all, abs, round, map, filter.
• Loops, wall-clock time, and result size are capped — keep programs bounded; project fields to stay under the result cap.
• Must assign result.
• HTTP errors surface the NetBox response body, including per-field validation errors — so a failed create()/update() returns actionable detail you can correct from and retry in the next program.

Injected helpers

The exact set is rendered live in the netbox_execute description and gated by capability (write/IPAM-allocation/GraphQL/scripts/plugins/products). Object types use dot notation ('dcim.device'); the path form ('dcim/devices') also resolves. Full catalog in [references/code-mode-helpers.md](references/code-mode-helpers.md) — load it when you need a specific signature. The essentials:

get(type, filters={}, fields=[], limit=50, offset=0, ordering=None) → {count, next, previous, results[]}
get_by_id(type, id, fields=[]) → object
search(query, object_types=[], limit=5) → {type: [results]}
options(type) → field schema (required fields, choices)
discover_models(app=None) / inspect_model(app, endpoint)        # enumerate types incl. plugins
create(type, data) / update(type, id, data) / delete(type, id)  # WRITE-gated
bulk_create / bulk_update / bulk_delete                          # WRITE-gated
available_prefixes / available_ips / allocate_prefix / allocate_ip ...   # IPAM-gated
graphql(query, variables=None, analyze=True)                     # GRAPHQL-gated
api(method, endpoint, params=None, data=None)                    # raw escape hatch

Filter rules (two footguns)

1. Never use in. NetBox silently ignores fieldin on many fields and returns all rows — confidently wrong. The helpers reject it. Pass a list as the field value instead:

``python # WRONG get('dcim.site', filters={'id__in': [1, 2, 3]}) # RIGHT — list value sends repeated query params, which NetBox honors get('dcim.site', filters={'id': [1, 2, 3]}) ``

1. No multi-hop traversal. device__site_id-style nested filters are rejected. Use a two-step query — fetch the parent's ids, then filter the child:

``python racks = get('dcim.rack', filters={'site': 'nyc-dc1'}, fields=['id']) rackids = [r['id'] for r in racks['results']] devices = get('dcim.device', filters={'rackid': rack_ids}, fields=['id', 'name']) result = devices['results'] ``

When to use rediscover_netbox()

Only when you have a specific reason to believe the tool inventory is stale: the user just installed/uninstalled a plugin, a tool you genuinely expected returns "tool not found," or a tool you no longer expect is still advertised. It is also the recovery path after a failed plugin registration. Do NOT call it before every operation, to "refresh" NetBox data (helpers are always live), or "just to be sure" at session start — each call costs a round-trip plus re-registration.

More end-to-end programs in [references/code-mode-workflows.md](references/code-mode-workflows.md).

Discrete Mode

When Code Mode isn't available, every operation is its own MCP tool with a typed signature and a rich docstring. Discrete mode delivers more context directly (the tool list and docstrings are in front of you), so there is less to discover — but each operation is a separate round-trip. Core tools:

Discrete-mode rules the docstrings already enforce (follow them):

• Always pass fields= to minimize tokens (80–90% payload reduction on reads). limit default 50, max 1000.
• The __in and multi-hop rules are identical to code mode — pass a list value; use the two-step pattern.
• netboxgetchangelogs: action must be imperative singular ("create"/"update"/"delete" — NetBox rejects "updated" with 400). changedobjecttype is "app.model"; changedobjecttype_id is the integer ContentType PK.
• Object types accept both "dcim.device" and "dcim/devices". Use discover_models() to enumerate types (including plugins).

Product & Plugin Namespaces

When the tenant is entitled and the integration is configured, additional helpers/tools appear:

• Plugins: branching (branch), changes (changerequest, reviews, comments), custom-objects (customtype/customobject_).
• Products: Assurance (assurance) and Discovery (discovery).

These are entitlement-gated — present only when your tenant has the product/plugin. Don't assume they exist; discover them. For their domain logic, see [netbox-branching](../netbox-branching/SKILL.md), [netbox-changes](../netbox-changes/SKILL.md), [netbox-custom-objects](../netbox-custom-objects/SKILL.md), [netbox-assurance](../netbox-assurance/SKILL.md), and [netbox-discovery](../netbox-discovery/SKILL.md).

Anti-Patterns

1. Writing import in code mode → "Forbidden operation: Import." The dominant failure. Use namespace helpers + allowed builtins only.
2. Not searching the schema first → guessing a wrong object type or unsupported filter. Confirm via netboxsearchschema before executing.
3. The __in footgun → silently-broad results / rejection. Use a list as the field value.
4. Multi-hop filters → rejected. Use the two-step pattern.
5. Forgetting to assign result → wasted turn.
6. One netbox_execute per object instead of a single looped/bulk program → defeats the point of code mode.
7. Over-calling rediscover_netbox as a precaution → it's a round-trip cost, not a data refresh.
8. Not projecting fields= → oversized payloads, hitting the result cap, wasted tokens (both modes).
9. Assuming a product/plugin namespace exists → it's entitlement-gated. Discover it.
10. NameError on an undefined name (ip_network, re, …) → only injected helpers + allowed builtins exist.

References

• [references/code-mode-helpers.md](references/code-mode-helpers.md) — Full injected-helper catalog with signatures, grouped by capability gate. Load when you need an exact signature.
• [references/code-mode-workflows.md](references/code-mode-workflows.md) — End-to-end one-program examples (reads, list-of-ids, two-step joins, IPAM allocation, changelog audits).