There are two MCP servers for NetBox from NetBox Labs itself, plus four more from outside the company — and they differ far more than the shared name suggests. The open source one is read-only with four tools; the managed one is a NetBox Cloud service with nearly a hundred, and it is the only option that routes agent writes through a branch a human has to merge. This page lists what exists as of 15 August 2026, what each one covers, where your credentials sit, and what happens when an agent writes.
This page puts all six MCP options we could find for NetBox side by side across ten dimensions. We make one of them, which is why the sourcing rules below matter: no scores, no stars, no rankings — every cell states a fact and links to the evidence.
The real choice is not the tool count but the shape of access: a small NetBox-specific agent connector, NetBox Cloud’s integrated agentic change workflow, or one governance layer across NetBox and further systems. As of 15 August 2026 the field looks like this — six MCP options, plus seven adjacent projects that turn up under the same search terms but are not comparable options. Every claim in the table is a summary of sourced cells in the comparison matrix below.
Self-hosted NetBox alongside other systems, behind one governance layer
Yes — every tool carries an access class (read/write/dangerous/admin) that policies can gate
A standing service to operate; no /api/plugins/ (NetBox Branching), no GraphQL, no bulk writes
Hands-on tested
What separates them. Coverage runs from 4 tools to 608. Only the two NetBox Labs servers discover plugin object types at runtime. Only three of the six give callers their own identity — the rest work through a single NetBox token, and the OSS server’s HTTP transport is unauthenticated unless you set MCP_AUTH_TOKEN, which its README says plainly. On write safety the spread is widest: three options write straight to the live instance, nbox splits writes into plan and apply, and only the Platform server isolates them in a branch with a diff a human approves.
A Rust binary that is a terminal UI, a CLI and an MCP server at once. The MCP side is read-first: eleven read tools plus every object as an nbox:// resource, with two opt-in write tools behind a plan/apply step. The only community option with OIDC and a per-user vault on its HTTP write path. Self-declared pre-1.0.
The managed first-party server, in Public Preview on NetBox Cloud since June 2026. Nearly 100 tools covering reads, writes, bulk operations, IPAM allocation, cable tracing, GraphQL, config rendering and scripts, plus other NetBox Labs products. It is the only option here that routes agent writes through isolated branches a human reviews and merges, and the only one with dynamic model discovery that picks up installed plugins. Nothing to install; NetBox Cloud only.
A write-enabled derivative of the NetBox Labs OSS server: its four read tools plus more than fifty named write tools for sites, tenants, tags, VLANs, IPAM, DCIM, circuits and virtualization, and three generic create/update/delete tools. Ships a Docker Compose path and HTTP transport for web clients. The README states plainly that plugin object types are not supported. No license is declared.
The open source first-party server and, by adoption, the reference implementation: 213 stars and 94 forks, last pushed the day before this survey. Deliberately small -- four read tools over 127 core object types, no writes, no GraphQL -- and the project states that keeping it small is the point, with forking under Apache 2.0 as a supported path. The only option here with a signed, provenance-carrying container image.
Nine generic tools -- get, get-by-id, create, update, delete, the three bulk variants and changelogs -- that work against any NetBox object type rather than one tool per type. The most-starred community option with writes. The README calls it production ready while the roadmap still lists input validation, tests and async support as open. Created and last pushed on the same day in September 2025.
Not a NetBox-specific server but a self-hosted gateway reading a declarative YAML description of the NetBox REST API -- 608 tools, 98% of the API surface, the broadest coverage in this list. Governance (per-tool and per-user authorization, server-side credentials, output policies, audit) comes from the gateway rather than the connector. Two gaps follow from the static description: no dynamic plugin discovery, and no access to /api/plugins/ endpoints, which rules out driving NetBox Branching.
Rows are the ten dimensions; columns are the options. The table scrolls horizontally. Superscript numbers link to the source for that cell.
“Not documented” means the sources we checked do not state it — not that the feature is absent. If you maintain one of these projects and we got something wrong, open an issue and we will fix the data file.
Yes -- agents work in isolated branches, propose diffs and merge only through human review; change request workflows and NetBox Validation pre-change analysis are wired into the same connection (Professional+) 10
No -- NetBox Branching is a plugin served under /api/plugins/branching/, and those endpoints are explicitly out of scope for netbox.dadl, so branches cannot be driven through the gateway today 18
Dry-run / confirmation before a write
Yes -- --dry-run, or --allow-writes plus an explicit --confirm; the MCP side splits this into plan_write and apply_write 1
Branch diffs act as the review step; the docs additionally recommend enabling tool approval prompts in the MCP client 2
A gateway that reads a static description of the NetBox REST API reaches exactly what that description declares — no less, and structurally no more. Four things sit outside ours, and for those a purpose-built server wins:
/api/plugins/ endpoints — which is where NetBox Branching actually lives. The branching REST API sits under /api/plugins/branching/, and netbox.dadl explicitly excludes plugin endpoints, so branch-based review of agent writes cannot be driven through the gateway today. The Platform MCP Server has branching wired in natively.
Dynamic model discovery. A static description cannot notice at startup which plugins an instance has installed. The Platform MCP Server discovers every object type the instance exposes; the OSS server can optionally discover plugin object types on NetBox 4.2+.
GraphQL. netbox.dadl covers the REST API only. The Platform MCP Server offers GraphQL queries from the Starter tier.
Bulk collection operations. Bulk PUT/PATCH/DELETE on collections is explicitly not covered in netbox.dadl — per-object calls in Code Mode are the stated alternative. The Platform MCP Server and netbox-mcp-rw both ship dedicated bulk tools.
A gateway reading a static description reaches none of these. If your use case depends on one of them, pick accordingly.
A gateway earns its keep when NetBox is not the only system in the picture:
Questions that span systems. “Which hosts are critical in Checkmk, and who owns them in NetBox?” is one agent turn against one endpoint, instead of two connectors that cannot see each other. The public DADL registry lists 30 APIs and over 4,000 tools as of August 2026, so the second system is often already described.
One place for credentials. The NetBox token lives server-side and is injected at call time, rather than sitting in every operator’s client configuration.
One authorization model and one audit trail across every connected system, instead of per-connector conventions.
Context economy at scale. 608 NetBox tools is a lot of schema for a model to carry; a catalog that stays flat matters more the more systems you connect. This one is not ours alone — the Platform MCP Server’s Code Mode solves the same problem for NetBox Cloud.
The honest trade-off: a gateway is a standing service you have to run, monitor and keep available. The dedicated servers’ leanest shape is a subprocess your LLM client launches — nothing to operate at all. Run one of them centrally over HTTP instead and you are operating a component too; the count evens out, and the difference shifts to what the matrix above shows: a service fronting one system on a shared NetBox token versus a multi-user gateway fronting several. Either way, a gateway only pays for itself once more than one system is behind it. If NetBox is the only thing you will ever connect, and reads are all you need, the first-party open source server is the leaner answer. It exists, it is four tools, and it cannot change a single NetBox object.
The machine-readable coverage report for our option — every tool, its access class, and which parts of the NetBox REST API are and are not covered — is at dadl.ai/d/netbox. It records 608 tools over 608 of roughly 620 REST endpoints (98%), each tool carrying one of four access classes (248 read, 207 write, 114 dangerous, 39 admin), with the uncovered areas named explicitly: plugin endpoints, GraphQL and bulk collection writes.
Yes — two, and they barely overlap. The open source NetBox MCP Server (Apache-2.0) is deliberately small: four read tools over 127 core object types, no writes, no GraphQL, and by adoption it is the reference implementation. The Platform MCP Server is a managed NetBox Cloud service in Public Preview since June 2026: nearly 100 tools (the vendor's figure) covering reads, writes, bulk operations, GraphQL, config rendering and scripts, with agent writes routed through isolated branches. Only the first has a public repository — a GitHub search surfaces it, not the managed service.
Can an AI agent write to NetBox through MCP, or only read?
It depends on the option. The NetBox Labs OSS server is read-only by design — no write tools exist. nbox is read-only by default, with two opt-in write tools that are deliberately narrow and split into plan and apply. NetBox MCP Enhanced adds more than fifty named write tools plus three generic ones to the OSS server's four read tools. netbox-mcp-rw exposes generic create, update and delete including bulk variants. The Platform MCP Server writes from the Starter plan upward, individually or in bulk. ToolMesh + netbox.dadl declares write and delete tools across the API, each classed read, write, dangerous or admin so a policy can gate them.
Which option can use NetBox Branching, so agent changes are reviewed before they merge?
Only the Platform MCP Server. It works in isolated branches, proposes diffs, and merges only through human review, with change request workflows and NetBox Validation pre-change analysis on the same connection. NetBox Branching is a plugin whose REST API lives under /api/plugins/branching/, and netbox.dadl explicitly excludes plugin endpoints — so branches cannot be driven through our gateway today. The remaining write-capable options write directly to the live instance; nbox's plan/apply step is a confirmation gate, not an isolated branch.
Where do my NetBox credentials end up?
nbox reads the API token from a config.toml (0600 on Unix, redacted in config show) or an environment variable. NetBox MCP Enhanced and the NetBox Labs OSS server take it from the environment of the machine running the server; netbox-mcp-rw reads it from the LLM client's configuration or environment on the client machine. The Platform MCP Server takes a NetBox v2 token as a header on every connection, so the secret sits in each caller's client configuration and the agent acts under that token's RBAC permissions. ToolMesh keeps the token server-side in the gateway's credential store and injects it at call time, so it never appears in a client configuration.
Do these work with a self-hosted NetBox behind a firewall?
Five of the six do, because you run them yourself inside your network. The exception is the Platform MCP Server: it is part of NetBox Cloud and works only there, with NetBox Enterprise support stated as planned for later this year.
Do plugin object types work over MCP?
Only through the two NetBox Labs servers. The Platform MCP Server discovers every object type the instance exposes at startup, plugins included. The OSS server has optional plugin discovery (ENABLE_PLUGIN_DISCOVERY, NetBox 4.2+, off by default), which extends its type registry. NetBox MCP Enhanced states plainly that plugin object types will not work, netbox-mcp-rw lists plugin support as an open roadmap item, and netbox.dadl excludes /api/plugins/ endpoints as out of scope for a general description.
Reference from here on — the adjacent projects, the methodology and disclosure, the changelog and the sources. Nothing below is needed for the decision above; it is where the evidence lives.
An Apache-2.0 library of Agent Skills from NetBox Labs that teach an agent NetBox conventions, API patterns and branching workflows. Not an MCP server -- it is knowledge, not access -- but it is installed alongside one, and it works with any of the options here.
An AI assistant from NetBox Labs embedded in the NetBox UI, in public preview. It is used inside NetBox, not called by an external agent, so it is not an MCP server.
A small FastMCP server over streaming HTTP with four tools covering sites and site groups, written up in a blog series and therefore easy to find. Its README states it is provided as-is and no license is declared, and the tool surface is too narrow to compare across the ten dimensions.
A write-enabled fork of the NetBox Labs OSS server that tracks upstream automatically, Apache-2.0 and actively pushed. Listed as adjacent rather than as an option because it is a fork whose stated purpose is to follow another listed option; worth promoting to a full entry if it develops an independent scope.
A NetBox MCP server in Go under AGPL-3.0, created in August 2026. Too new to assess -- one week of history at the time of this survey. Re-check at the next registry diff.
An Apache-2.0 demonstration server from a network automation training channel, created in January 2025. The repository carries no README documenting a tool surface, so nothing beyond dimension 1 could be filled.
A Docker suite bundling MCP servers for seven systems, NetBox among them. It is a multi-system bundle rather than a dedicated NetBox server, which is a different category from every option on this page.
Disclosure. This page is published by Dunkel Cloud GmbH, which makes ToolMesh — one of the six options listed. We have an obvious interest in how it looks. The countermeasures are structural: options are ordered alphabetically by display name, our own entry included; every cell carries a source link; no cell contains a score or a ranking; and the two “when to choose” sections above name the cases where our option is the wrong tool.
Inclusion criterion. A GitHub search for “netbox mcp”, “netbox mcp server” and “netbox model context protocol” on 15 August 2026 returned 46 repositories; the managed Platform MCP Server has no public repository and entered through the accompanying vendor research. An option is listed when all three points hold: a publicly reachable source or product page; presented as a usable server rather than a demo, course artifact or personal fork without independent substance; and documentation sufficient to fill at least the profile dimension. Six options met all three — five public repositories plus the managed Platform MCP Server. Seven projects are documented as adjacent, and the rest are forks, course artifacts and one-day experiments without a README or license.
How the facts were gathered. Each option’s own repository, README, documentation and — where cited — source code, read on 15 August 2026. Repository metadata — license, last commit, stars and forks — comes from the GitHub API on the same date. Where a project does not state something, the cell says “not documented” rather than guessing. For hosted products that means the platform documentation, not only a product page: the Platform MCP Server’s capabilities are documented across the NetBox Cloud documentation and the vendor’s launch post, and both are cited. Registry aggregators were used only to find candidates, never as evidence.
Depth of verification. The five options we do not maintain are documentation and source-code review, not test results — read, not run. The verification column in the decision table states this per option, and vendor figures we cannot count ourselves are quoted as such. Our own entry is the exception: the netbox.dadl behind it is live-verified, and its row was last reviewed on 15 August 2026. The maintainers of all listed projects were notified when this page went live; if a documentation review got something wrong, their corrections land in the data file.
Corrections and additions. The matrix lives as one YAML file in this site’s repository; the site repository itself is private, so corrections travel as issues on the public ToolMesh repository. Correcting a fact, or adding an option we missed, takes one issue. Maintainers: if we have misrepresented your project, that is a bug and we will fix it.
First publication. Six MCP options compared across the ten dimensions, plus seven adjacent projects found under the same search terms that are not comparable options. Facts taken from each project's own repository and documentation; our own entry is backed by the live-verified netbox.dadl.