Porter hosts an MCP server at https://mcp.porter.run. Point an MCP client at that URL, sign in, and your agent can do the same things you’d do in the dashboard or the CLI: read logs, check why a deploy failed, change a service, redeploy it.
You don’t need to install or run anything locally. The server runs on Porter’s side, so all your client needs is the URL and a browser to sign in.
Connect your client
Quick install
For the CLI clients, one command is enough:
Client configuration
Claude Code
Codex
OpenCode
Cursor
Claude Desktop
-
Add the server:
Or add it to
.mcp.json to share it with the project:
-
Start Claude Code and run
/mcp.
-
Select porter and sign in through the browser.
For more information, see the Claude Code MCP documentation.
-
Add the server:
Or add it to
~/.codex/config.toml directly:
-
Sign in:
The ChatGPT desktop app, Codex CLI, and IDE extension share this file, so you only do this once.For more information, see the Codex MCP documentation.
-
Run the setup command:
-
Enter
porter as the name, choose remote, and paste the URL:
Or add it to opencode.json directly:
-
Sign in:
OpenCode also starts the sign-in on its own the first time you use a Porter tool.For more information, see the OpenCode MCP documentation.
-
Create or edit
.cursor/mcp.json in your project root:
Use ~/.cursor/mcp.json instead to get Porter in every project.
-
Restart Cursor.
-
Open Settings > Tools & MCP, select porter, and click Connect.
For more information, see the Cursor MCP documentation.
-
Go to Customize → Connectors, click +, then Add custom connector. On Team or Enterprise, an Owner does this from Organization settings → Connectors.
-
Paste the URL:
-
Click Add and approve access in the browser.
Porter then shows up under Connectors in the + menu of any conversation. On Team or Enterprise, everyone else finds it already listed and clicks Connect.Claude connects to remote MCP servers from Anthropic’s cloud, not from your machine. That’s true in the desktop app too.
Capabilities
You ask for things in plain language, so it’s easier to think in terms of tasks than tools:
- Deploy an app. Create an application and get it running without leaving your agent. If you’re deploying a pre-built image, the agent can take it all the way. If you’re deploying from a GitHub repo, the agent creates the app and opens a pull request with the build workflow, and you merge it to deploy your code.
- Update an app. Change services, resources, or domains on an existing app and redeploy. Updates can be validated with a dry run first, so you can see the resulting config before anything deploys.
- Troubleshoot your app. Read logs, metrics, and notifications, or check nodes, node groups, pods, and load balancers to find what’s broken. From there you can go back to the config that caused it, fix it, and redeploy.
- Rightsize a service. Compare requested CPU and memory against what the service actually uses, then adjust the requests.
- Copy an environment. Create dev or staging versions of an existing app, or move one to a different cluster.
- Inspect resources. List projects, apps, and clusters, or check the status of a deployment.
Some examples:
The tools reference has the full list of tools.
Limitations
OAuth is currently the only auth method, so headless environments like CI pipelines and sandboxed agents are not supported.
Risks
MCP tools can perform destructive operations on your applications, including
creating, updating, and redeploying services. They can directly impact
production environments and cause downtime to services if not used with care.
- Use non-production environments first. Validate changes in a dev or staging project before applying them to production.
- Review the agent’s plan before approving it, particularly for operations that modify production resources.
- Disable auto-approval against production. It is reasonable on a test project and unsafe everywhere else.
- Account for prompt injection. The agent reads your application logs, notifications, and pull request bodies, all of which can carry attacker-controlled text. The server cannot distinguish injected instructions from your own, so a compromised agent can call
create_app, update_application, or redeploy_application and change or restart a running service. Read tools exclude environment variables and secrets, and redact token-shaped values in logs, so the exposure is to your application’s configuration and availability rather than its credentials.
Revoke access
You revoke access per client. Remove the stored credentials, then delete the server entry so it doesn’t sign in again later.
Claude Code
Codex
OpenCode
Cursor
Claude Desktop
logout clears the stored OAuth credentials and remove deletes the server entry.That clears the stored OAuth tokens. Also remove the porter entry from opencode.json to get rid of the server entirely. Delete the porter entry from .cursor/mcp.json (or ~/.cursor/mcp.json) and restart Cursor.
Go to Customize → Connectors, click the three dots next to Porter, and choose Remove. On Team and Enterprise plans, an Owner has to remove it for the whole organization from Organization settings → Connectors.
