docs: added MCP docs

Author: evilsocket

docs/index.md

@@ -10,6 +10,7 @@ Nerve is an ADK ( _Agent Development Kit_ ) with a convenient command line tool
   - [Record & Replay](#-record--replay)
   - [Adding Tools](#️-adding-tools)️
   - [Conversation Window](#-conversation-window)
+* [Model Context Protocol](mcp.md)
 * [Namespaces](namespaces.md)
 * [Workflows](workflows.md)
 
@@ -277,24 +278,7 @@ nerve play agent-trace.jsonl -f # much faster
 
 ### 🛠️ Adding Tools
 
-Since version 1.5.0 Nerve is integrated with [Model Context Protocol (MCP)](https://modelcontextprotocol.io/introduction). Any server from [the multitude of publicly available MPC servers](https://github.com/punkpeye/awesome-mcp-servers) can be used with:
-
-```yaml
-agent: You are a helpful assistant.
-task: Write something to your knowledge graph, then read it back, save it to output.txt and set your task as complete.
-
-using:
-  - task
-
-mcp:
-  memory:
-    command: npx
-    args: ["-y", "@modelcontextprotocol/server-memory"]
-
-  filesystem:
-    command: npx
-    args: ["-y", "@modelcontextprotocol/server-filesystem", "."]
-```
+Since version 1.5.0 Nerve is integrated with [Model Context Protocol (MCP)](https://modelcontextprotocol.io/introduction), you can refer to the [MCP section](mcp.md) of the documentation to use MCP tools.
 
 When a tool can be represented as a shell command, you can conveniently extend the agent capabilites in the YAML:
 

docs/mcp.md

@@ -0,0 +1,90 @@
+# Model Context Protocol
+
+Nerve supports [MCP](https://modelcontextprotocol.io/introduction) both as a client and as a server:
+
+* [Client](#mcp-client)
+* [Server](#mcp-server)
+    - [Serving an Agent](#serving-an-agent)
+    - [Serving an Agent and its Tools](#serving-an-agent-and-its-tools)
+    - [Tools Only](#tools-only)
+
+## MCP Client
+
+An agent can use any server from [the multitude of publicly available MPC servers](https://github.com/punkpeye/awesome-mcp-servers) with:
+
+```yaml
+agent: You are a helpful assistant.
+task: Write something to your knowledge graph, then read it back, save it to output.txt and set your task as complete.
+
+using:
+  - task
+
+mcp:
+  memory:
+    command: npx
+    args: ["-y", "@modelcontextprotocol/server-memory"]
+
+  filesystem:
+    command: npx
+    args: ["-y", "@modelcontextprotocol/server-filesystem", "."]
+```
+
+## MCP Server
+
+Nerve can also act as a MCP server via the `nerve serve <agent>` command. In this mode you can turn an agent into a tool for other agents to use.
+
+### Serving an Agent
+
+For instance this command line will serve the [code-audit](https://github.com/evilsocket/code-audit) agent as an MCP server:
+
+```bash
+nerve serve code-audit --mcp
+```
+
+This means that agents can use other agents as tools, for instance:
+
+```yaml
+agent: You are a helpful assistant.
+task: Perform a code audit of {{ path }}.
+
+using:
+  - task
+
+mcp:
+  code_audit:
+    command: nerve
+    args: ["serve", "code-audit", "--mcp"]
+```
+
+You can find a full [example here](https://github.com/evilsocket/nerve/tree/main/examples/mcp-recipe).
+
+### Serving an Agent and its Tools
+
+Additionally you can also serve the agent tools:
+
+```bash
+nerve serve code-audit --mcp -t
+```
+
+This will export both the agent itself and its tools via MCP.
+
+### Tools Only
+
+You can use this mechanism to serve simple tools via MCP as well. For instance if you create a `tools.yml` like this:
+
+```yaml
+tools:
+  - name: get_weather
+    description: Get the current weather in a given place.
+    arguments:
+        - name: place
+          description: The place to get the weather of.
+          example: Rome
+    tool: curl wttr.in/{{ place }}
+```
+
+You can serve it via MCP with:
+
+```bash
+nerve serve tools.yml --mcp
+```