cml-mcp-server
# cml-mcp-server
An MCP (Model Context Protocol) server for Cisco Modeling Labs (CML) v2.10.
It wraps Cisco's official `virl2_client` Python SDK to expose lab/topology
CRUD, node/link management, node configuration, and a catalog of ready-made
example network architectures as MCP tools - so an MCP client (Claude Code,
Claude Desktop, etc.) can build and run network labs for testing designs and
configs.
## Setup
```bash
cd cml-mcp-server
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
cp cml-config.example.toml cml-config.toml # then fill in host / username / password
chmod 600 cml-config.toml
```
### Credentials: config file vs. env vars
Connection settings can come from either a TOML config file or environment
variables; env vars win if both are present. The config file is the
recommended approach since it keeps credentials out of shell history and
process-list-visible env, and out of MCP client config JSON.
`cml-config.toml` (project root) - searched first - or
`~/.config/cml-mcp-server/config.toml`, or an explicit path via
`CML_CONFIG_PATH`:
```toml
[cml]
host = "cml.example.com"
username = "api-user"
password = "change-me"
# verify_ssl = false
```
Equivalent env vars: `CML_HOST`, `CML_USERNAME`, `CML_PASSWORD`,
`CML_VERIFY_SSL`. `CML_VERIFY_SSL` accepts `true`/`false` or a path to a CA
bundle; it defaults to `false` since most CML controllers run with a
self-signed certificate.
## Running
```bash
source .venv/bin/activate
cml-mcp-server
```
This starts the server on stdio, ready to be attached to an MCP client.
### Claude Code / Claude Desktop config
```json
{
"mcpServers": {
"cisco-cml": {
"command": "/absolute/path/to/cml-mcp-server/.venv/bin/cml-mcp-server"
}
}
}
```
Or point at an explicit config file / use env vars instead:
```json
{
"mcpServers": {
"cisco-cml": {
"command": "/absolute/path/to/cml-mcp-server/.venv/bin/cml-mcp-server",
"env": {
"CML_CONFIG_PATH": "/absolute/path/to/cml-config.toml"
}
}
}
}
```
## Tools
**Labs**
- `list_labs`, `get_lab_topology`
- `create_lab`, `delete_lab`
- `start_lab`, `stop_lab`, `wipe_lab`
**Platform catalog**
- `list_node_definitions` - available platform types (iosv, iosvl2,
csr1000v, nxosv, asav, alpine, unmanaged_switch, ...), depends on which
reference images are imported into your CML controller
**Nodes**
- `add_node`, `remove_node`
- `start_node`, `stop_node`
- `get_node_config`, `set_node_config` (day-0 startup config)
- `extract_node_config` (pull the running config off a booted node)
- `get_node_console_log`
**Links**
- `add_link`, `remove_link` - by default connects the next free interface
on each node; interface labels can be given explicitly
**Bulk topology building**
- `build_topology` - create a whole lab (nodes + links + configs) in one
call from a node/link spec
- `list_example_topologies`, `get_example_topology`,
`deploy_example_topology` - a small catalog of ready-made architectures:
- `two_router_ospf` - single-area OSPF adjacency
- `bgp_two_as_peering` - eBGP between two ASes
- `three_tier_campus` - core/distribution/access skeleton
- `leaf_spine_datacenter` - 2-spine/2-leaf full-mesh underlay
- `firewall_dmz` - IOS zone-based firewall with inside/outside/DMZ zones
These templates default to `iosv`/`iosvl2` (bundled with every CML
install); swap `node_definition` per node for platform-specific images
(e.g. `nxosv`, `csr1000v`, `asav`) if you have them.
## Typical flow
1. `list_node_definitions()` - see what platforms are available.
2. Either:
- `list_example_topologies()` → `deploy_example_topology("two_router_ospf")`, or
- describe an architecture and let the assistant design a node/link spec,
then call `build_topology(...)` directly.
3. `start_lab(lab_id)`, then `get_lab_topology(lab_id)` to check node state.
4. `get_node_console_log` / `extract_node_config` to inspect a running node;
`set_node_config` + `wipe_lab` + `start_lab` to iterate on day-0 config.
5. `delete_lab(lab_id)` when done - this is permanent.
## Security notes
- This server can create, start, and delete labs and nodes on the target
CML controller. Scope the CML account to a dedicated user if you don't
want it operating with full admin rights, and treat `delete_lab`,
`wipe_lab`, and `remove_node` as destructive, unrecoverable actions.
- Credentials are only ever read (from `cml-config.toml` or env vars);
nothing is written to disk by this server. Keep the config file out of
version control (it's in `.gitignore` by default) and permissioned `600`.
- Session/token handling (login, refresh, re-auth) is delegated entirely to
`virl2_client`.
TDQS
Scored across 22 tools
Each tool targets a distinct resource and action, from lab-level operations (create_lab, start_lab) to node-level (add_node, stop_node) and link-level (add_link, remove_link). Aggregates like build_topology and example-specific tools are clearly separated, and even similar-looking config tools (get_node_config vs extract_node_config) are explicitly differentiated.
All tools follow a consistent verb_noun pattern in snake_case (e.g., list_labs, create_lab, add_node, delete_lab). Verbs and nouns are predictable, and related actions use clear prefixes like start_/stop_, add_/remove_, and get_/list_.
With 22 tools, the server sits in the heavy range (16-25) according to the calibration rubric. While each tool serves a clear purpose, the sheer number might be overwhelming for an agent, and some operations (e.g., example topology deployment) could have been consolidated without losing clarity.
The tool surface covers core lab lifecycle (create, delete, start, stop, wipe), node management (add, remove, config, console), and link management (add, remove). However, there is no direct update for lab metadata or node resources (RAM, image), and no way to edit a link without recreating it, which are minor gaps agents can work around.