Runtime MCP server
Run your agents from the AI app you already use. One URL connects Claude, ChatGPT, Grok Bot, Cursor, VS Code and Codex to Runtime.
Claude
ChatGPT
Grok Bot
Cursor
VS CodeCodexEverything Runtime, from your chat
Everything Runtime, from your chat
Connect your AI app
Paste one URL into Claude, ChatGPT, Grok Bot or Cursor. Sign in with your company SSO, pick a workspace and allow. Your agents are a message away.
Add custom connector
Name
Runtime
Remote MCP server URL
https://mcp.runtm.com/mcp
Signed in to Runtime with Okta as [email protected]Connect Claude
Claude wants to use Runtime as you.
Runtime connectedYour agents, sessions and approvalsAgents and triggers
Ask which agents you run and what wakes them up. Point a Slack channel, a Teams chat, an inbox or a schedule at one without leaving the conversation.
Which agent handles partner-bank requests? Send the new bank inbox to it.
Checking your Runtime agents.
Payment Ops handles partner-bank requests. Mail sent to [email protected] now goes straight to it.
Sessions
Start a run on its own computer and follow it as it works. Ask a follow-up, or have it run a query inside the sandbox.
Re-review Harbor Vape Supply against our VAMP thresholds.
Starting a run on the Underwriting agent.
The dispute ratio is 1.4% against a 0.9% threshold. Underwriting recommends a 25% rolling reserve.
Approvals
When an agent needs a person, the request comes to you. See what it wants to do and why, then approve or deny it in chat.
Is anything waiting on me?
Looking for approvals.
Payment Ops wants to release a $48,200 payout to Acme Ltd. The invoice matches and it's within policy.
Skills and runbooks
Hand over a runbook and it becomes a skill on the right template. Every agent on that template follows it from the next run.
Add our merchant re-review runbook to the underwriting template.
Saving it as a skill.
Done. Underwriting agents follow merchant-re-review v1.0.0 from their next run.
Guardrails
Say in plain words what agents must never do alone. Runtime enforces it on every run, and only admins can change it.
Make every refund over $5,000 wait for a person.
Adding an org-wide guardrail.
Saved. Refunds over $5,000 now pause for an approver, for every agent in the org.
Results
Ask how an agent did this week: what it handled, how its runs were graded, and what it cost. Then fix what it got wrong.
How did Support do this week?
Pulling this week's runs.
212 runs and 94% passed, at $0.41 a run. Most misses were refunds without an order ID. Want me to update the skill to ask for one first?
Set up your client
- Chat apps
- Claude.ai, Claude Desktop, ChatGPT and Grok Bot add the URL and sign in to Runtime.
- IDEs and terminals
- Claude Code, Cursor, VS Code and Codex sign in the same way, or send an API key.
- Nothing to run
- A remote server over streamable HTTP. No local process, no Docker.
Run this in your terminal. Then type /mcp in Claude Code, pick runtime and sign in to Runtime.
claude mcp add --transport http runtime https://mcp.runtm.com/mcpIn Claude.ai or Claude Desktop, open Settings → Connectors → Add custom connector. Paste this URL, then sign in to Runtime.
https://mcp.runtm.com/mcpIn ChatGPT, turn on Developer mode under Settings → Apps & Connectors → Advanced settings, then click Create. Paste this URL and sign in to Runtime.
https://mcp.runtm.com/mcpSend this message to your bot and confirm. Then click Authorize on the card it posts and sign in to Runtime.
Add a custom MCP server called runtime at https://mcp.runtm.com/mcpAdd this to ~/.cursor/mcp.json. Then click Connect next to runtime in Cursor Settings → MCP.
{
"mcpServers": {
"runtime": { "url": "https://mcp.runtm.com/mcp" }
}
}Add this to .vscode/mcp.json. Then click Start above the server and sign in to Runtime.
{
"servers": {
"runtime": { "type": "http", "url": "https://mcp.runtm.com/mcp" }
}
}Run these in your terminal. The second command opens your browser to sign in to Runtime.
codex mcp add runtime --url https://mcp.runtm.com/mcp
codex mcp login runtimeSecure by default, SSO ready
Connect with the sign-in your team already uses. Every token acts for one person in one workspace, with the same permissions they have in Runtime.
SAML · OIDCYour SSO, not another password
Claude, ChatGPT and your IDE sign in through Runtime with your company SSO: Okta or Duo, over SAML or OIDC. Pick a workspace, approve, and you are connected.
Tokens that act for a person
Each token belongs to one person in one workspace and only works on the MCP server. Roles and rate limits apply on every call, and a connected app can't create API keys.
Access that follows your team
Remove someone from the workspace and their connection stops on the next call. Tools a person has no permission for are hidden from their AI app.
Credentials stay out of chat
Steps that need a secret return a dashboard link instead. A 403 stays a 403: the agent names the missing permission and stops.
Your agents' own MCP servers can sit behind your identity provider too. Give the server an OAuth app, with the Okta preset or any OAuth issuer, and agents sign in to it. Okta setup guide
MCP questions
What is the Runtime MCP server URL?+
https://mcp.runtm.com/mcp. It is a remote MCP server over streamable HTTP, so there is nothing to install or run locally.
How do I sign in?+
Claude.ai, Claude Desktop, ChatGPT and Grok Bot sign in with OAuth: paste the URL, sign in to Runtime with your company SSO, Google or email, pick a workspace and approve. Claude Code, Cursor, VS Code and Codex can also send an API key as a Bearer token. Create an API key in Runtime under Settings > API Keys and export it as RUNTM_API_KEY.
Does it work with our SSO?+
Yes. People sign in to Runtime through Okta or Duo over SAML or OIDC, and connecting an AI app uses that same sign-in. Every token is scoped to one person and one workspace.
Which tools does it expose?+
About 30 tools by default across agents, sessions, context, guardrails and triggers, plus docs search and a generic call for anything else in the Runtime API. Add ?toolsets=all to the URL for file and lifecycle tools, or ?mode=tools for one tool per API operation. Tools your permissions can't use are hidden.
Can our agents use MCP servers behind Okta?+
Yes. An org admin gives the MCP server its own OAuth app with the Okta preset: your Okta domain, authorization server ID, and the scopes openid profile email offline_access. Agents then sign in to that server without dynamic client registration.
How is the MCP server different from the CLI?+
Both drive the same Runtime API with the same permissions. The CLI is for terminals and scripts; the MCP server is for chat apps and IDEs, and for people who never open a terminal.
Build your team of agents
Start from chat, the terminal or the dashboard. Same agents, same guardrails, same audit trail.