TL;DR:

• Claude Managed Agents lets you run cloud-based AI agents with MCP tools like Firecrawl and Linear, no local setup required.
• The four core concepts are agents (model + tools + prompt), environments (cloud containers), sessions (running agent instances), and events (messages in and out).
• Skipping the quick-start wizard and doing manual setup gives you a clear mental model, making it much easier to debug and extend your agent.

Claude's cloud console now includes Managed Agents: a way to run AI agents entirely in the cloud, connected to external tools via MCP. In this post I'll walk through the full manual setup, connecting Firecrawl for web search and Linear for task creation, so you end up with an agent that automatically fetches AI news and logs it as issues in your Linear project.

Claude Managed Agents is a feature inside the Claude cloud console (currently in beta) that lets you deploy and run AI agents without managing any local infrastructure. Instead of running Claude Desktop on your machine, you configure agents that live in the cloud, call MCP tools, and operate inside isolated compute environments called sessions.

The system is built around four core concepts: agents, environments, sessions, and events. Understanding what each one does before touching any configuration makes the whole setup significantly easier to reason about.

Before touching any configuration, it helps to have a clear model of how the pieces relate.

Agents define what Claude does: the model, system prompt, tools, MCP servers, and skills. You create an agent once and reference it by ID across as many sessions as you like.

Environments are configured container templates: pre-installed packages (Python, Node.js, Go, etc.), network access rules, and mounted files. Think of an environment as the compute context your sessions run inside. By default, outbound connections to external MCP server URLs are blocked, which is something you will need to change.

Sessions are running agent instances. A session references an agent and an environment, performs a specific task, and generates outputs. Multiple sessions can run inside the same environment independently.

Events are the messages exchanged between your application and the agent: user turns, tool results, and status updates. When you send a message to a session, you are sending an event. The agent streams responses back as server-sent events, and the full event history is persisted server-side.

Credential vaults are a separate UI concept specific to the console. They store API keys and OAuth tokens independently of any agent or environment, so the same vault can be reused across projects. Crucially, the LLM never sees your credentials directly: it only calls MCP tools that execute using those credentials behind the scenes.

Also read: MCP vs. A2A Protocols: What Developers Need to Know About AI's New Plumbing

Firecrawl MCP is a remote MCP server that exposes Firecrawl's web scraping and search capabilities as tools any MCP-compatible agent can call. Instead of writing code to hit the Firecrawl API, your agent calls tools like firecrawl_search and gets back clean markdown that it can immediately reason over.

The remote MCP URL for Firecrawl is:

https://mcp.firecrawl.dev/v2/mcp

This is the URL you will use when configuring the credential vault and the agent JSON.

Linear is a project management tool built for software teams. The Linear MCP server exposes Linear's API as callable tools, letting an agent create issues, update task statuses, and interact with your Linear workspace without writing any integration code.

Authentication for Linear uses OAuth rather than a bearer token, so the setup flow is slightly different: you select Linear from the credential vault options and follow the OAuth prompt to connect your account.

Start here, even though vaults appear last in the sidebar. Having credentials ready before you configure your agent and environment keeps the setup linear (no pun intended).

1. In the Claude cloud console, go to Credential Vaults and click New Vault.
2. Give it a name. In the video the vault is called MCP vault, grouping all MCP credentials for this project in one place.
3. Click Create, then click Add Credential inside the new vault.
4. For Firecrawl: select Bearer Token, choose Custom Server, and enter https://mcp.firecrawl.dev/v2/mcp as the server URL. Then paste your Firecrawl API key. You can create a key from the Firecrawl dashboard under API Keys.
5. Check the acknowledgment box and click Add Credential. Status should show as Active.
6. Click Add Credential again, this time select Linear. Follow the OAuth flow to connect your Linear workspace.

Both credentials should now show as active inside your vault.

1. Go to Environments and click Add Environment.
2. Give it a name (for example my environment). The only available hosting type is cloud.
3. Click Create.

At this point the environment is created with default settings. You can leave everything as-is for now, but you will need to come back to toggle one network setting before your session can reach external MCP servers. See the troubleshooting section below.

Go to Agents and click New Agent. A guided builder opens that tries to configure your agent using an LLM, but during beta it does not always handle custom MCP servers correctly. If you describe your agent and hit Generate, you may get a config that is missing the Firecrawl MCP server and its tools entirely. The more reliable path is to work directly in JSON.

You can use the Templates section as a starting point. The JSON structure consists of two key arrays:

• mcp_servers: one object per MCP you want to use, with a name, url, and type of "url"
• tools: one object per tool you want available, with a name and a policy

The policy field controls execution behavior. Set it to "auto" to let the agent call the tool without asking, or "confirmation" if you want it to pause and check with you first. For an automated pipeline, use "auto".

A practical shortcut: grab the JSON from an existing template in the console, paste it into Claude, and ask it to add the specific MCP server you need with the correct structure. This avoids writing it from scratch and catches formatting mistakes before you try to save.

After clicking Create Agent, you should see Firecrawl and Linear listed under the agent's connected tools, alongside your system prompt and model. If anything looks wrong, you can go back into the JSON at any time to fix it.

1. Go to Sessions and click Create Session.
2. Select your agent (for example AI news to Linear agent).
3. Select your environment.
4. Under Credential Vaults, add the vault you created. Once added, you will see the MCP servers identified from the agent config. They will show a Connected via MCP Vaults label once credentials are matched.
5. Check the I own or am authorized to use this vault box.
6. Give the session a name (optional) and click Create Session.

If you see a console error about the MCP server host being blocked, it means the environment's network policy is preventing outbound connections to your MCP server URLs.

To fix it:

1. Cancel the session creation.
2. Go back to Environments, click into your environment, and click Edit.
3. Toggle on Allow MCP server network access.
4. Click Save Changes.
5. Create the session again.

This is the most common blocker when setting up custom MCP servers and is easy to miss because the default is locked down.

Once your session is created, send it a message:

Fetch recent AI news for me and create tasks based on the three most important stories in Linear.

You can monitor execution two ways:

• Transcript: shows what the agent is communicating to the user, higher level and readable
• Debug: shows each individual tool call, the parameters sent, and the responses received

In the debug view you will see firecrawl_search trigger with a limit of 10, searching for recent AI news. After results come back, the agent uses the Linear MCP to create issues for the top stories. Each issue includes the title and key details from the article.

Note: If the agent searches for "latest AI news in 2025" instead of the current year, update your system prompt to include the current date. Building agents is iterative: check the debug output, refine the prompt, and run again.

If you need to change the agent's MCP servers or system prompt after creating it, do not use the Ask Claude input at the top of the console. Instead:

1. Go to the Agents tab.
2. Select your agent.
3. Click the three-dot menu (not the Edit button) and select Guided Edit.

This opens the same JSON editor alongside a help panel, giving you full control over the configuration without losing your existing setup.

If a session has been idle and you want to resume it:

1. Go to the session view.
2. Click Actions in the top right.
3. Select Send Event.
4. Type your message in the input field that appears.

This is the way to interact with a session that is paused or waiting for input without creating a new session from scratch.

The Firecrawl plus Linear combination is just one example. The same architecture works for any workflow that chains data retrieval with structured output:

• Pull competitor product updates from the web and log them as issues in your backlog
• Monitor a set of URLs for content changes and create tasks when something new appears
• Aggregate job postings from multiple sources and triage them into a Linear project
• Fetch research papers on a topic and create a summary issue for each
• Track mentions of your product across the web and surface them to your team

Also read: Context Layer for AI Agents: Can You Automate Context Feeding into Your Agents?

The managed agent architecture is especially well-suited to pipelines that need to pull from one service and push structured output to another. Credential vaults keep your keys secure, environments isolate your compute, and the session model makes it easy to run multiple independent workflows side by side.

Check out How to Schedule Recurring Web Tasks with Claude Desktop and Firecrawl for a similar automation pattern running locally on Claude Desktop. If you want to go deeper on the MCP protocol itself, read our MCP vs. A2A explainer.