docs(examples): add CrewAI MCP example

mcp/examples/README.md

@@ -5,6 +5,7 @@ End-to-end demos showing how to wire the e2a MCP surface into popular agent fram
 | Framework | Path | LLM | Stdio variant | Hosted variant |
 | --- | --- | --- | --- | --- |
 | LangChain (LangGraph ReAct) | [langchain/](./langchain/) | Anthropic Claude | `agent.py` | `agent_hosted.py` |
+| CrewAI | [crewai/](./crewai/) | Anthropic Claude | `agent.py` | `agent_hosted.py` |
 | Google ADK | [adk/](./adk/) | Google Gemini | `agent.py` | `agent_hosted.py` |
 | OpenAI Agents SDK | [openai-agents/](./openai-agents/) | OpenAI GPT | `agent.py` | `agent_hosted.py` |
 | OpenAI Codex CLI | [codex/](./codex/) | Codex (OpenAI) | `[mcp_servers.e2a]` in `config.toml` | `[mcp_servers.e2a-hosted]` in `config.toml` |

mcp/examples/crewai/README.md

@@ -0,0 +1,45 @@
+# CrewAI × e2a
+
+Single-agent CrewAI crew that drives the e2a [MCP server](https://www.npmjs.com/package/@e2a/mcp-server) via [`crewai-tools`](https://github.com/crewAIInc/crewAI-tools)' `MCPServerAdapter`. Picks up the e2a tool surface and uses it to answer natural-language email tasks.
+
+Two transport options:
+
+- **`agent.py`** — runs the MCP server locally via `npx -y @e2a/mcp-server` (stdio). Simplest for laptop dev; needs a Node toolchain.
+- **`agent_hosted.py`** — talks to the hosted endpoint at `https://mcp.e2a.dev/mcp` (Streamable HTTP). Pick this when deploying to serverless runtimes (Cloud Run, Lambda) where spawning a stdio child process is awkward or impossible, or when you don't want a Node toolchain on the agent host.
+
+## Prerequisites
+
+- Python 3.10+
+- An [e2a API key](https://e2a.dev)
+- An [Anthropic API key](https://console.anthropic.com/)
+- For `agent.py` only: Node 18+ (the script shells out to `npx -y @e2a/mcp-server`)
+
+## Run (local stdio)
+
+```bash
+pip install -r requirements.txt
+export E2A_API_KEY=e2a_…
+export E2A_AGENT_EMAIL=your-bot@your-domain.com   # optional default inbox
+export ANTHROPIC_API_KEY=sk-ant-…
+
+python agent.py "what's in my inbox?"
+python agent.py "reply to the most recent message politely"
+```
+
+## Run (hosted)
+
+```bash
+pip install -r requirements.txt
+export E2A_API_KEY=e2a_…
+export ANTHROPIC_API_KEY=sk-ant-…
+
+python agent_hosted.py "what's in my inbox?"
+```
+
+If your account has exactly one agent, the hosted endpoint auto-resolves it at session init — no `E2A_AGENT_EMAIL` needed. With multiple agents, pass `agent_email` per tool call.
+
+## How it works
+
+`MCPServerAdapter` connects to either a stdio child process (via `StdioServerParameters`) or a Streamable HTTP endpoint (via a dict with `transport: "streamable-http"` and a Bearer header). Inside the `with` block, the adapter yields one CrewAI tool per MCP tool — wired straight into the `Agent` so the crew can call them.
+
+To swap models, change `"anthropic/claude-sonnet-4-6"` to any [LiteLLM-compatible model string](https://docs.litellm.ai/docs/providers) CrewAI accepts (CrewAI uses LiteLLM under the hood).

mcp/examples/crewai/agent.py

@@ -0,0 +1,79 @@
+"""End-to-end demo: CrewAI agent driving @e2a/mcp-server over stdio.
+
+Wires the e2a MCP server into a single-agent CrewAI crew so the LLM can
+send, read, and reply to email through natural-language prompts.
+
+Requires:
+  E2A_API_KEY        e2a API key (https://e2a.dev)
+  E2A_AGENT_EMAIL    (optional) default agent inbox
+  E2A_BASE_URL       (optional) self-hosted e2a base URL
+  ANTHROPIC_API_KEY  Anthropic API key
+
+Run:
+  pip install -r requirements.txt
+  python agent.py "what's in my inbox?"
+"""
+
+import os
+import sys
+
+from crewai import Agent, Crew, Process, Task
+from crewai_tools import MCPServerAdapter
+from mcp import StdioServerParameters
+
+BACKSTORY = (
+    "You manage email through the e2a tools. Call whoami once to find "
+    "your inbox address. Use list_messages and get_message to read; "
+    "use reply_to_message (not send_email) when replying to an existing "
+    "thread so In-Reply-To and References headers are preserved."
+)
+
+
+def _e2a_env() -> dict[str, str]:
+    env = {"E2A_API_KEY": os.environ["E2A_API_KEY"]}
+    for k in ("E2A_AGENT_EMAIL", "E2A_BASE_URL"):
+        if k in os.environ:
+            env[k] = os.environ[k]
+    return env
+
+
+def main(prompt: str) -> None:
+    server_params = StdioServerParameters(
+        command="npx",
+        args=["-y", "@e2a/mcp-server"],
+        env=_e2a_env(),
+    )
+
+    with MCPServerAdapter(server_params) as e2a_tools:
+        print(
+            f"Loaded {len(e2a_tools)} e2a tools: "
+            f"{', '.join(t.name for t in e2a_tools)}\n"
+        )
+
+        agent = Agent(
+            role="Email Manager",
+            goal="Handle the operator's email request precisely and concisely.",
+            backstory=BACKSTORY,
+            tools=e2a_tools,
+            llm="anthropic/claude-sonnet-4-6",
+            allow_delegation=False,
+            verbose=True,
+        )
+        task = Task(
+            description=prompt,
+            expected_output="A clear, concise answer to the user's email-related request.",
+            agent=agent,
+        )
+        crew = Crew(
+            agents=[agent],
+            tasks=[task],
+            process=Process.sequential,
+            verbose=False,
+        )
+        result = crew.kickoff()
+        print(result)
+
+
+if __name__ == "__main__":
+    prompt = " ".join(sys.argv[1:]) or "what's in my inbox?"
+    main(prompt)

mcp/examples/crewai/agent_hosted.py

@@ -0,0 +1,76 @@
+"""End-to-end demo: CrewAI agent against the hosted e2a MCP server.
+
+Same shape as `agent.py` but uses the hosted MCP endpoint
+(https://mcp.e2a.dev/mcp) over Streamable HTTP instead of spawning
+@e2a/mcp-server locally over stdio. Pick this variant when:
+- Deploying the CrewAI agent to a serverless runtime (Cloud Run,
+  Lambda, etc.) where launching a stdio child process per request is
+  awkward or impossible.
+- You don't want a Node toolchain on the agent host.
+- You want updates to land without rebuilding the agent's image.
+
+Requires:
+  E2A_API_KEY        e2a API key (https://e2a.dev)
+  ANTHROPIC_API_KEY  Anthropic API key
+
+Run:
+  pip install -r requirements.txt
+  python agent_hosted.py "what's in my inbox?"
+"""
+
+import os
+import sys
+
+from crewai import Agent, Crew, Process, Task
+from crewai_tools import MCPServerAdapter
+
+BACKSTORY = (
+    "You manage email through the e2a tools. Call whoami once to find "
+    "your inbox address. Use list_messages and get_message to read; "
+    "use reply_to_message (not send_email) when replying to an existing "
+    "thread so In-Reply-To and References headers are preserved."
+)
+
+
+def main(prompt: str) -> None:
+    server_params = {
+        "url": "https://mcp.e2a.dev/mcp",
+        "transport": "streamable-http",
+        "headers": {
+            "Authorization": f"Bearer {os.environ['E2A_API_KEY']}",
+        },
+    }
+
+    with MCPServerAdapter(server_params) as e2a_tools:
+        print(
+            f"Loaded {len(e2a_tools)} e2a tools: "
+            f"{', '.join(t.name for t in e2a_tools)}\n"
+        )
+
+        agent = Agent(
+            role="Email Manager",
+            goal="Handle the operator's email request precisely and concisely.",
+            backstory=BACKSTORY,
+            tools=e2a_tools,
+            llm="anthropic/claude-sonnet-4-6",
+            allow_delegation=False,
+            verbose=True,
+        )
+        task = Task(
+            description=prompt,
+            expected_output="A clear, concise answer to the user's email-related request.",
+            agent=agent,
+        )
+        crew = Crew(
+            agents=[agent],
+            tasks=[task],
+            process=Process.sequential,
+            verbose=False,
+        )
+        result = crew.kickoff()
+        print(result)
+
+
+if __name__ == "__main__":
+    prompt = " ".join(sys.argv[1:]) or "what's in my inbox?"
+    main(prompt)

mcp/examples/crewai/requirements.txt

@@ -0,0 +1,3 @@
+crewai[anthropic]>=0.80.0
+crewai-tools[mcp]>=0.25.0
+email-validator>=2.0