rcrewai 0.2.1 → 0.4.0

data/bin/rcrewai

@@ -1,7 +1,7 @@
 #!/usr/bin/env ruby
 # frozen_string_literal: true
 
-require "bundler/setup"
-require "rcrewai"
+require 'bundler/setup'
+require 'rcrewai'
 
-RCrewAI::CLI.start(ARGV)
+RCrewAI::CLI.start(ARGV)

data/docs/mcp.md

@@ -0,0 +1,109 @@
+# MCP (Model Context Protocol) Support
+
+RCrewAI 0.3 ships a minimal MCP client that lets your agents call tools
+hosted by any MCP server — local processes (stdio) or remote services
+(streamable HTTP) — without any custom adapter code.
+
+## What is MCP?
+
+[MCP](https://modelcontextprotocol.io) is an open protocol for connecting
+language models to tools and data sources. Servers expose tools via a
+JSON-RPC schema; clients (like RCrewAI) connect, discover the tools, and
+invoke them on the model's behalf.
+
+The big win: any of the dozens of off-the-shelf MCP servers — filesystem,
+git, Slack, Postgres, browser automation, custom internal tools — works
+with RCrewAI without writing a single line of glue.
+
+## Connecting
+
+### Stdio (local subprocess)
+
+```ruby
+require 'rcrewai'
+
+RCrewAI::MCP::Client.with_connection(
+  command: "npx",
+  args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
+) do |client|
+  puts "connected to #{client.server_name}"
+  client.tools.each { |t| puts " - #{t.name}: #{t.description}" }
+end
+```
+
+### Streamable HTTP (remote)
+
+```ruby
+RCrewAI::MCP::Client.with_connection(
+  url: "https://mcp.example.com/v1",
+  headers: { "Authorization" => "Bearer #{ENV['MCP_TOKEN']}" }
+) do |client|
+  # ...
+end
+```
+
+### Manual lifecycle
+
+`with_connection` is the recommended form (auto-closes on block exit), but
+you can also drive the lifecycle manually:
+
+```ruby
+client = RCrewAI::MCP::Client.connect(command: "ruby", args: ["my_server.rb"])
+# ... use client.tools ...
+client.close
+```
+
+## Tool name prefix
+
+To prevent collisions when an agent uses tools from multiple MCP servers,
+tool names are prefixed with the server name:
+
+```
+echo-server__echo
+filesystem__read_file
+filesystem__write_file
+git__commit
+```
+
+The prefix is `#{server_name}__#{tool_name}`, where `server_name` comes
+from the server's own `serverInfo.name`.
+
+## Using MCP tools with an agent
+
+MCP tools are ordinary `RCrewAI::Tools::Base` instances — pass them to
+`Agent.new(tools: ...)` like any other tool:
+
+```ruby
+RCrewAI::MCP::Client.with_connection(command: "npx", args: [...]) do |client|
+  agent = RCrewAI::Agent.new(
+    name: "fs_agent",
+    role: "Filesystem operator",
+    goal: "Read and summarize files",
+    tools: client.tools
+  )
+
+  task = RCrewAI::Task.new(name: "summarize",
+                           description: "Read /tmp/notes.md and summarize",
+                           agent: agent)
+  result = agent.execute_task(task)
+  puts result[:content]
+end
+```
+
+When `ToolRunner` (the new native-function-calling loop) is selected, the
+LLM picks an MCP tool by name, RCrewAI dispatches the call to the MCP
+server, threads the result back into the conversation, and continues.
+
+## What's not (yet) supported in 0.3
+
+The 0.3 client is intentionally minimal — it covers `initialize`,
+`tools/list`, `tools/call`, and the `notifications/initialized` handshake.
+The following are not implemented:
+
+- **Resources** (`resources/list`, `resources/read`)
+- **Prompts** (`prompts/list`, `prompts/get`)
+- **Server mode** — RCrewAI can be an MCP client, not yet an MCP server
+- **OAuth** authentication for HTTP transports
+- **Notifications other than `initialized`** (e.g., progress, log)
+
+These will land in follow-up releases as the MCP spec stabilizes.