Skip to content

Integrations

Nautobot Agent includes built-in chart and Mermaid skills, optional custom skills, bundled native model integrations, and optional Model Context Protocol (MCP) servers. See Configuration for the complete settings reference and provider examples.

Custom skills

Custom skills are global instructions sent to the configured model provider. They do not grant Nautobot permissions or add tools. Do not include secrets or instructions that imply broader authorization than users already have.

  1. Confirm the intended administrators have the native Skill permissions; see Permissions.
  2. Open Nautobot Agent → Skills and add a name, description, Markdown instructions, and enabled state.
  3. Review the full text before enabling it; enabled custom instructions are included in execution.
  4. Confirm through normal operation that data and tool access remain governed by Nautobot permissions.

The packaged nautobot-charts and nautobot-mermaid skills are included by default and cannot be edited or disabled in the Skills interface. Administrators can exclude both by setting PLUGINS_CONFIG["nautobot_agent"]["builtin_skills"] = False; see Configuration. Enabled custom skills remain available, but cannot use the reserved built-in identifiers. Disabling packaged skills removes their instructions, not chart or Mermaid rendering capabilities.

LLM configuration and credential profile

Only superusers can save LLM configurations because provider options determine where deployment credentials are sent.

  1. Put provider credentials in a named credentials profile in deployment configuration, using the deployment's secret-management mechanism. Do not put credentials in the model's Parameters field.
  2. In Nautobot Agent → LLM Configurations, enter provider:model, choose the matching Credential profile, and enter non-secret provider options in Parameters as a JSON object.
  3. Assign exactly one configuration as the default Agent model. A separate scope-classifier configuration is optional; without one, the Agent model is used for classification.
  4. Ensure the Agent model supports tool calling and exposes a positive integer profile.max_input_tokens. Missing context metadata prevents Agent tools from running. Classifier-only configurations are exempt.
  5. Save, then reload the Agent page. Reload web and worker processes after deployment configuration changes.

Profile values take precedence over matching Parameters values. Model IDs and saved configurations do not guarantee that a provider/model combination is available or compatible. See LLM configurations and model roles.

MCP servers

Only enable MCP servers that the Nautobot operator trusts. MCP tools can access systems or cause external side effects using the configured credentials and worker identity. Every user who can chat can call every enabled server; there are no per-user or per-group MCP access restrictions. Nautobot permissions and blocked_models do not restrict external MCP servers. Review its operator, available tools, credentials, worker identity, network access, and possible effects before enabling it.

  1. Define required profiles in PLUGINS_CONFIG["nautobot_agent"]: use mcp_credentials for HTTP header values and mcp_environments for stdio environment variables. Store secrets in deployment secret management.
  2. Ensure the endpoint or executable is available to every worker that can run chat, then reload processes after deployment configuration changes.
  3. In Nautobot Agent → MCP Servers, enter a descriptive name, select the transport, provide its settings, select the applicable profile if needed, and leave Enabled off during review.
  4. Test from the worker environment and review the server's tools and effects. Enable it only after review; repeat the review when its endpoint, executable, profile, or worker identity changes.

MCP records contain profile names and connection or launch settings, not deployment secrets. Every value in a selected profile must be set, even when the server is disabled. A server's name is shown to chat users when the integration is unavailable and sent to the model provider; use a non-sensitive label.

HTTP transport

  1. Select HTTP and provide the server's HTTP(S) URL. It must include a hostname and must not include user information or a fragment. Do not put credentials in the URL or query string.
  2. Leave the stdio launch field empty.
  3. If authentication headers are needed, select a named mcp_credentials profile.
  4. Confirm that workers can reach the endpoint and that its TLS and network access meet deployment requirements.

Standard input/output (stdio) transport

  1. Select Standard input/output and leave the HTTP URL empty.
  2. Enter a JSON launch object with a nonblank command and optional args (string array), cwd (absolute path), and env (string-to-string object). Do not include an mcpServers wrapper or type/transport keys. For example:
{"command": "runner", "args": ["--serve"]}
  1. Install the executable and its dependencies on every eligible worker. It runs with the Nautobot worker account's filesystem and network access; saving the form does not check that the executable exists.
  2. If needed, select an mcp_environments profile. Profile values override matching values in launch env; keep secrets in the deployment profile.
  3. Test the command as the worker account before enabling the server.

For a Python-packaged server, install it into a dedicated environment accessible to the worker. Example for the reference time server:

python -m venv /opt/mcp/time-venv
/opt/mcp/time-venv/bin/pip install mcp-server-time
/opt/mcp/time-venv/bin/python -m mcp_server_time --help

Example launch JSON on Linux:

{"command": "/opt/mcp/time-venv/bin/python", "args": ["-m", "mcp_server_time"]}

On Windows, use forward slashes or escape backslashes as \\ in JSON:

{"command": "C:/mcp/time-venv/Scripts/python.exe", "args": ["-m", "mcp_server_time"]}

Use the interpreter where the package is installed. Paths are resolved on the worker, not the browser's computer.

If the form reports a missing command, provide the bare launch object above rather than client configuration wrapped in mcpServers. For invalid JSON on Windows, use forward slashes or double backslashes. If a saved server is unavailable, check that the command and package are accessible to the worker and that any cwd exists on that worker.

A lost response, timeout, or cleanup warning does not establish whether an external operation happened or was rolled back. Nautobot Agent does not automatically retry MCP operations. Reconcile the external system before retrying an interrupted or uncertain operation; see Operations.