|
| 1 | +--- |
| 2 | +title: Run vMCP locally with the CLI |
| 3 | +sidebar_label: Local CLI mode |
| 4 | +description: |
| 5 | + Run a Virtual MCP Server on your workstation with thv vmcp, without |
| 6 | + Kubernetes. |
| 7 | +--- |
| 8 | + |
| 9 | +Most of the Virtual MCP Server (vMCP) guides deploy vMCP to Kubernetes using |
| 10 | +the operator. The ToolHive CLI also ships a local mode for the same vMCP |
| 11 | +runtime, so you can aggregate MCP servers from a ToolHive group on your |
| 12 | +workstation without a cluster. |
| 13 | + |
| 14 | +Local mode is useful for: |
| 15 | + |
| 16 | +- **Prototyping aggregation** against a handful of MCP servers before |
| 17 | + committing to a `VirtualMCPServer` resource in Kubernetes. |
| 18 | +- **Local development** against the same optimizer, conflict resolution, and |
| 19 | + composite-tool logic that the operator uses. |
| 20 | +- **Demos** where you want a single endpoint fronting several MCP servers |
| 21 | + without the operator overhead. |
| 22 | + |
| 23 | +The CLI command is `thv vmcp` with three subcommands: `serve`, `init`, and |
| 24 | +`validate`. They share the `pkg/vmcp/` runtime with the Kubernetes operator, |
| 25 | +so configuration, tool aggregation, and the optimizer behave the same way in |
| 26 | +both environments. |
| 27 | + |
| 28 | +## Prerequisites |
| 29 | + |
| 30 | +- ToolHive CLI `v0.24.0` or later. See |
| 31 | + [Install the ToolHive CLI](../guides-cli/install.mdx). |
| 32 | +- One or more MCP servers running in a ToolHive group. See |
| 33 | + [Run MCP servers](../guides-cli/run-mcp-servers.mdx) and |
| 34 | + [Group management](../guides-cli/group-management.mdx) to create a group |
| 35 | + and add servers to it. |
| 36 | +- Docker or Podman if you plan to enable the Tier 2 semantic optimizer, which |
| 37 | + starts a Text Embeddings Inference (TEI) container. |
| 38 | + |
| 39 | +## Quick mode: aggregate a group in one command |
| 40 | + |
| 41 | +When you already have a ToolHive group with running MCP servers, the fastest |
| 42 | +way to start vMCP is quick mode. Pass `--group` and vMCP generates a minimal |
| 43 | +in-memory configuration from the group at startup: |
| 44 | + |
| 45 | +```bash |
| 46 | +thv vmcp serve --group demo-tools |
| 47 | +``` |
| 48 | + |
| 49 | +This is equivalent to creating a config file with: |
| 50 | + |
| 51 | +- `groupRef: demo-tools` |
| 52 | +- `incomingAuth.type: anonymous` |
| 53 | +- `outgoingAuth.source: inline` |
| 54 | +- `aggregation.conflictResolution: prefix` with prefix format `{workload}_` |
| 55 | + |
| 56 | +vMCP binds to `127.0.0.1:4483` and aggregates every accessible MCP server in |
| 57 | +the group behind that single endpoint. |
| 58 | + |
| 59 | +:::warning[Quick mode is loopback-only] |
| 60 | + |
| 61 | +Because quick mode uses anonymous authentication, `thv vmcp serve --group` |
| 62 | +rejects non-loopback bind addresses. Valid values for `--host` are |
| 63 | +`127.0.0.1`, `localhost`, or any other loopback address. To expose vMCP on |
| 64 | +another interface or add real authentication, switch to a config file (see |
| 65 | +below). |
| 66 | + |
| 67 | +::: |
| 68 | + |
| 69 | +### Connect a client |
| 70 | + |
| 71 | +In another terminal, register the vMCP endpoint as a remote MCP server so your |
| 72 | +configured AI clients can discover it: |
| 73 | + |
| 74 | +```bash |
| 75 | +thv run http://localhost:4483/mcp --name local-vmcp |
| 76 | +``` |
| 77 | + |
| 78 | +If you haven't registered any AI clients yet, run `thv client setup` first. |
| 79 | +See [Client configuration](../guides-cli/client-configuration.mdx) for |
| 80 | +details. |
| 81 | + |
| 82 | +## Config-file mode: generate, review, serve |
| 83 | + |
| 84 | +When you need incoming OIDC auth, a non-loopback bind, per-backend outgoing |
| 85 | +auth, or composite tools, use a configuration file. The `init` subcommand |
| 86 | +scaffolds one from an existing group, `validate` checks it, and `serve` |
| 87 | +runs it. |
| 88 | + |
| 89 | +### Step 1: generate a starter config |
| 90 | + |
| 91 | +`thv vmcp init` enumerates the workloads in a group and emits a YAML config |
| 92 | +with one backend entry per accessible server: |
| 93 | + |
| 94 | +```bash |
| 95 | +thv vmcp init --group demo-tools --output vmcp.yaml |
| 96 | +``` |
| 97 | + |
| 98 | +The `--group` flag is required. If you omit `--output`, the generated YAML |
| 99 | +is written to stdout so you can pipe it through `less` or redirect it |
| 100 | +yourself. |
| 101 | + |
| 102 | +A generated file looks similar to this (comments trimmed): |
| 103 | + |
| 104 | +```yaml title="vmcp.yaml" |
| 105 | +name: demo-tools-vmcp |
| 106 | +groupRef: demo-tools |
| 107 | + |
| 108 | +incomingAuth: |
| 109 | + type: anonymous |
| 110 | + |
| 111 | +outgoingAuth: |
| 112 | + source: inline |
| 113 | + |
| 114 | +aggregation: |
| 115 | + conflictResolution: prefix |
| 116 | + conflictResolutionConfig: |
| 117 | + prefixFormat: "{workload}_" |
| 118 | + |
| 119 | +backends: |
| 120 | + - name: fetch |
| 121 | + url: http://localhost:24162/mcp |
| 122 | + transport: streamable-http |
| 123 | + - name: osv |
| 124 | + url: http://localhost:24163/mcp |
| 125 | + transport: streamable-http |
| 126 | +``` |
| 127 | +
|
| 128 | +Edit the file to switch to OIDC, change the conflict-resolution strategy, add |
| 129 | +composite tool definitions, or enable the optimizer. See |
| 130 | +[Configure vMCP servers](./configuration.mdx) for the complete schema. |
| 131 | +
|
| 132 | +### Step 2: validate the config |
| 133 | +
|
| 134 | +Before starting the server, run the validator to catch syntax errors and |
| 135 | +schema violations: |
| 136 | +
|
| 137 | +```bash |
| 138 | +thv vmcp validate --config vmcp.yaml |
| 139 | +``` |
| 140 | + |
| 141 | +A valid config exits with status `0` and no output. An invalid config exits |
| 142 | +non-zero with a descriptive error, for example: |
| 143 | + |
| 144 | +```text |
| 145 | +Error: validation failed: aggregation.conflictResolution: unknown strategy "preffix" |
| 146 | +``` |
| 147 | + |
| 148 | +### Step 3: start the server |
| 149 | + |
| 150 | +Once the config validates, start vMCP: |
| 151 | + |
| 152 | +```bash |
| 153 | +thv vmcp serve --config vmcp.yaml |
| 154 | +``` |
| 155 | + |
| 156 | +When `--config` is set, `--group` is ignored. Pass `--host` and `--port` to |
| 157 | +override the default `127.0.0.1:4483` bind address (non-loopback addresses |
| 158 | +are allowed in config-file mode). |
| 159 | + |
| 160 | +## Enable the optimizer |
| 161 | + |
| 162 | +vMCP supports the same [tool optimizer](./optimizer.mdx) in local mode as in |
| 163 | +Kubernetes. The CLI exposes three tiers through flags on `thv vmcp serve`. |
| 164 | + |
| 165 | +| Tier | Flag | What it does | External service | |
| 166 | +| ---- | ----------------------- | ---------------------------------------------------------------------------------------------------------- | ----------------------- | |
| 167 | +| 0 | _(none)_ | Pass-through: backends' tools are exposed as-is. | None | |
| 168 | +| 1 | `--optimizer` | FTS5 keyword optimizer: clients see only `find_tool` and `call_tool`, backed by an in-process SQLite FTS5 index. | None | |
| 169 | +| 2 | `--optimizer-embedding` | Tier 1 plus TEI semantic search. Implies `--optimizer`. | Managed TEI container | |
| 170 | + |
| 171 | +Tier 2 starts a managed TEI container on first run and stops it when the |
| 172 | +server exits. Customize the model and image with `--embedding-model` and |
| 173 | +`--embedding-image`: |
| 174 | + |
| 175 | +```bash |
| 176 | +thv vmcp serve --group demo-tools \ |
| 177 | + --optimizer-embedding \ |
| 178 | + --embedding-model BAAI/bge-small-en-v1.5 \ |
| 179 | + --embedding-image ghcr.io/huggingface/text-embeddings-inference:cpu-latest |
| 180 | +``` |
| 181 | + |
| 182 | +The defaults use the upstream CPU image; switch to a GPU image if you have |
| 183 | +one available. First-start container pulls can take 30-60 seconds. |
| 184 | + |
| 185 | +## Enable audit logging |
| 186 | + |
| 187 | +Pass `--enable-audit` to log each incoming request with the default audit |
| 188 | +configuration. If your config file already defines an `audit` section, the |
| 189 | +flag has no effect and the file's configuration wins. |
| 190 | + |
| 191 | +```bash |
| 192 | +thv vmcp serve --config vmcp.yaml --enable-audit |
| 193 | +``` |
| 194 | + |
| 195 | +See [Audit logging](./audit-logging.mdx) for the event format and how to |
| 196 | +customize it from a config file. |
| 197 | + |
| 198 | +## Local CLI vs Kubernetes deployment |
| 199 | + |
| 200 | +| Aspect | `thv vmcp serve` | `VirtualMCPServer` CRD | |
| 201 | +| --------------------- | ----------------------------------------------- | ---------------------------------------- | |
| 202 | +| Where it runs | Your workstation | Kubernetes cluster | |
| 203 | +| Backend discovery | Local ToolHive group | MCPGroup with MCPServer/MCPRemoteProxy | |
| 204 | +| Default bind | `127.0.0.1:4483` | Service in the cluster | |
| 205 | +| Authentication | Anonymous (quick mode) or as configured | Full OIDC, token exchange, embedded auth | |
| 206 | +| Lifecycle | Foreground process; Ctrl-C stops | Operator-managed Deployment | |
| 207 | +| Optimizer | Flag-driven (Tier 0/1/2) | `embeddingServerRef` on the CRD | |
| 208 | + |
| 209 | +Both paths load the same vMCP runtime from `pkg/vmcp/`, so a config file |
| 210 | +that validates locally will behave the same way when the operator loads it |
| 211 | +as a ConfigMap. This makes local mode useful for iterating on aggregation, |
| 212 | +conflict resolution, and composite-tool definitions before moving to |
| 213 | +Kubernetes. |
| 214 | + |
| 215 | +## Next steps |
| 216 | + |
| 217 | +- [Configure vMCP servers](./configuration.mdx) for the full config schema. |
| 218 | +- [Optimize tool discovery](./optimizer.mdx) for details on Tier 1 and |
| 219 | + Tier 2. |
| 220 | +- [Run vMCP in Kubernetes](./quickstart.mdx) when you're ready to move from a |
| 221 | + local process to the operator. |
| 222 | + |
| 223 | +## Related information |
| 224 | + |
| 225 | +- [`thv vmcp` command reference](../reference/cli/thv_vmcp.md) |
| 226 | +- [`thv vmcp serve`](../reference/cli/thv_vmcp_serve.md), |
| 227 | + [`thv vmcp init`](../reference/cli/thv_vmcp_init.md), |
| 228 | + [`thv vmcp validate`](../reference/cli/thv_vmcp_validate.md) |
| 229 | +- [Understanding Virtual MCP Server](../concepts/vmcp.mdx) |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4&size=40){kind=avatar}
0 commit comments