AI integrations
The NetConverter MCP server
Connect your AI assistant to the hosted MCP server, sign in with your NetConverter account, and understand view modes, write tools and limits.
Endpoint: https://api.netconverter.ai/mcp/v1/ · MCP Streamable HTTP · sign in with your NetConverter account. Keep the trailing slash.
What the MCP server does
The NetConverter MCP server lets your own AI assistant answer questions about your NetConverter jobs: interfaces, routes, NAT, rules, objects, flow simulation, validation evidence and the NetConverter knowledge base. It returns summaries, never raw device configuration. Most tools only read.
- Content: summaries and counts. Raw configuration is never returned, and credential values are redacted in every view.
- Tenant: taken only from your credential. You cannot ask for another tenant's data.
- No tool changes or deletes an existing job, and none connects to your devices. A flow simulation is a calculation over the parsed configuration.
- Every call is logged with the tool, account, job id and latency. Arguments are abstracted before they are stored.
The MCP tool reference lists every tool, generated from the tool catalog.
Connect your assistant
There are two connections. The hosted server at https://api.netconverter.ai/mcp/v1/ reviews existing jobs, findings and knowledge; it cannot read files from your computer. The local bridge in the NetConverter terminal submits a selected file by path. Choose your assistant explains both, and the per-client setup instructions cover ChatGPT, Claude and Claude Code step by step.
The transport is MCP Streamable HTTP, POST only. It is stateless, so there is no session to resume.
Sign in with your NetConverter account
The hosted server uses OAuth: your client opens the NetConverter sign-in page, you approve the consent screen, and every call is tied to you by name. No long-lived secret is stored on disk. Direct sign-in accepts known client applications only (Claude Code, Claude and ChatGPT). A client that registers itself automatically is refused with oauth_client_not_enrolled; connect it through the local bridge instead.
If your account belongs to more than one NetConverter organization, sign-in uses the organization you last had active in the NetConverter portal. If that is not one of your current organizations, the request is refused with ambiguous_org: switch organization in the portal and retry.
To cut access, an administrator removes the person from the organization. A token already verified can keep working for up to 30 seconds.
Quick accounts
Every Quick user gets MCP for their own jobs and the NetConverter knowledge base. Sign in with your NetConverter account; each user sees only their own jobs. Quick accounts use sign-in (OAuth) only: API keys are not available for MCP on Quick accounts.
Access on organization accounts
On a dedicated organization account, a NetConverter administrator enables MCP access for the organization or for named users. Without access, requests are refused with 403 mcp_not_granted. In a dedicated organization account, any user with MCP access can read every job in that organization. In a Quick account, each user sees only their own jobs. Teams that must be separated need separate organizations. To request access, ask your NetConverter contact.
View modes and privacy
The view mode is set on your MCP access grant; that is the ceiling. A session can lower it, never raise it. The set_view_level tool lowers the view for one MCP session for 24 hours. get_account reports the view in force.
| View | Addresses from your jobs | Names | Configuration-content tools | Credentials |
|---|---|---|---|---|
| Default | IP addresses and CIDRs from your own jobs are shown | Returned as written | Listed | Redacted |
| Strict | Replaced with <ip> and <cidr>, including addresses inside names | Tokenised | Hidden, and the write tools are hidden and refused | Redacted |
Knowledge-base reference text keeps its documentation-range example addresses in every view; it is product documentation, not your data. The local bridge has its own folder exposure level: its default, masked, asks the server for strict view and masks addresses again locally.
A tool result is placed into the conversation and sent to whichever model provider your client uses. NetConverter does not control that provider's retention or training terms; your agreement with the provider governs them. If addresses or names must not reach that provider, ask for strict view on your MCP access before you connect. Text in a configuration, such as a description, can appear in results; every result carries a notice telling the assistant to treat it as data. See privacy and data flow and the Trust and AI data-use statement.
Write tools
The write tools are off unless NetConverter enables them for your tenant, and they are never available in strict view. When enabled, apply_mapping_answers starts a new conversion of an existing job with your guided-mapping answers, propose_operations records decisions on a prescan, and submit_conversion queues a conversion. Each write tool records an audit entry before it runs; if it cannot, the call is refused and nothing changes. Use a client that asks you before it calls a tool.
Limits
500 tool calls and 5 USD of attributed cost per user per UTC day, a per-minute rate limit (120 for sign-in users by default), and a 64 KB cap on each response. The per-minute limit counts every request to the endpoint, including initialize and tools/list. Responses carry RateLimit-* headers, and get_account reports your effective limits.
What the server will not do
- Return configuration text, the full parsed model or a dump of any object.
- Accept a tenant or user identifier as an argument. Such arguments are ignored.
- Reveal whether a job id that is not yours exists. Unknown, foreign and unselected jobs all return the same message.
- Upload a configuration file. Use the local bridge for that.
Changes to the server and its tools are listed in the release notes.
Source and release context: Hosted MCP setup · Trust and AI data-use statement