Runtime

How to add Runtime to Claude.ai as a custom connector

In Claude, open Customize > Connectors, choose Add custom connector, enter https://api.withruntime.com/mcp, and approve in your browser.

With Runtime connected, Claude in the browser can create a Linux sandbox, run code in it and hand back files, with no API key pasted anywhere. The connector signs in through your browser and receives a key of its own, listed by name on your keys page and revocable in one click. New accounts get 50 free sandbox hours with no card, as of 25 September 2026 (trial).

Add the connector

On a Pro or Max plan:

  1. In Claude, go to Customize > Connectors.
  2. Choose +, then Add custom connector.
  3. Enter the remote MCP server URL: https://api.withruntime.com/mcp.
  4. Leave Advanced settings empty. Runtime's sign-in supports dynamic client registration, so it needs no client ID or secret.
  5. Choose Add.

Claude then opens a Runtime page in your browser:

  1. Sign in to Runtime, or create an account.
  2. The page shows which app is asking and where the key goes. Check it names Claude, then approve Connect.
  3. Back in Claude, every Runtime tool appears.

On a Team or Enterprise plan

Only an owner of the Claude organization adds a custom connector:

  1. Organization settings > Connectors, then Add.
  2. Hover over Custom and choose Web.
  3. Enter https://api.withruntime.com/mcp and choose Add.

Each member then opens Customize > Connectors, finds Runtime in the list and chooses Connect. Every member approves on Runtime with their own sign-in, so each gets a separate key that can be revoked alone. Invite them to your Runtime account first, so the connection lands in the shared account: see how to invite your team.

Use it in a chat

Turn the connector on from the + button at the lower left of the chat box, under Connectors. Then ask in plain words:

Create a Runtime sandbox, install pandas, load this CSV and plot revenue by month. Stop the sandbox when you are done.

Claude calls the tools itself. The ones it uses most:

Tool What Claude does with it
runtime_sandbox_create Starts a sandbox and waits until it is running
runtime_sandbox_exec Runs a command and reads the exit code and output
runtime_sandbox_files_write Puts your code or data into the sandbox
runtime_sandbox_files_read Reads results back
runtime_sandbox_interpreter_run A notebook-style cell; charts come back as images
runtime_sandbox_previews_create An HTTPS address for an app it started on a port
runtime_sandbox_manage Stops, pauses or wakes the sandbox
runtime_account_get Your trial time left and your credit
runtime_docs_read Any page of Runtime's docs, when it needs to check a detail

The server answers Claude's first contact with instructions: the quick start, how to retry safely and how to report a problem. That is why Claude knows to check exitCode and to stop a sandbox at the end.

How the connection works

Fact Detail
Endpoint https://api.withruntime.com/mcp, MCP Streamable HTTP
Sign-in OAuth 2.1 with PKCE and dynamic client registration
Where Claude calls from Anthropic's cloud, not your computer
Credential A key for a new agent named after the app
What it can do Manage resources and spend prepaid credit across the account
Revoke API keys, where it is listed by name
Free Claude plan One custom connector

Because Claude reaches the server from Anthropic's cloud, nothing is installed on your machine, and the connector works wherever you use the same Claude account.

Mistakes and how Runtime handles them

  • Approving a request you did not start. The approval page names the app; approve only when it is the Claude you just used.
  • Worrying about spend. Set a daily spending limit on the connector's row at API keys. Past it, a create or wake fails with spending_limit_reached and nothing is charged.
  • Leaving sandboxes running. Ask Claude to stop them, or let them pause: a sandbox made with default settings and never used pauses after 300 seconds and stops paying the running rate.
  • Tools with old names in a saved prompt. On 24 September 2026 every tool took its product's name, such as runtime_sandbox_exec. A call to an old name fails and names the new one.

Sources

Checked 25 September 2026.

Facts on this page were checked on 25 September 2026.