Gumloop MCP: Connect Claude, Cursor and Your Own MCP Servers

Gumloop MCP works in two directions. The Gumloop MCP server at mcp.gumloop.com/gumloop/mcp lets Claude, Cursor, ChatGPT, Codex and other MCP clients work with your Gumloop agents, sessions, skills and connected apps through 87 tools ^[1]. Going the other way, your agents can call outside MCP servers added as custom, proxied or hosted connectors. Both directions work on Pro, which is $37 a month for the whole organization's shared pool of 20,000 credits, and the server also works during the 14-day trial, which needs a card and turns into paid Pro unless you cancel before it ends. Setup takes one personal API key. Most custom-server failures trace back to a single fact: Gumloop connects from its own cloud, not from your laptop.

Server or client: decide which door you need

The naming confuses people, so separate the pieces first. The Gumloop MCP server is the door out: your coding assistant or chat app reaches into Gumloop. Custom, proxied and hosted MCP servers are doors in: a Gumloop agent reaches someone else's tools, next to its 300+ built-in connectors. A single agent can also be published on the MCP channel, which exposes that agent as an MCP server other AI tools can call. Gumstack is a fourth thing again, a gateway product sold on its own. Pick the direction first.

Connect Claude, Cursor or ChatGPT to the Gumloop MCP server

Start in Gumloop, not the client. Open Profile settings and generate a key in the Gumloop API key row; the server then handles the OAuth sign-in by itself. The documented clients are Claude on the web and desktop, Claude Code, Cursor, VS Code, Windsurf and the Gumloop CLI, and Gumloop's troubleshooting page adds ChatGPT and Codex. Clients that lack remote MCP support connect through mcp-remote. The steps below follow the docs for the most common setup, Claude and Claude Code.

1. Generate a personal Gumloop API key in Profile settings.
2. In Claude on the web or desktop, add a custom connector with the server URL and set Team to Personal.
3. In Claude Code, run claude mcp add --transport http gumloop followed by the server URL with its https prefix.
4. Copy the URL exactly, path included. The addresses mcp.gumloop.com/mcp and www.gumloop.com/mcp/ do not work as servers.
5. Ask the client to list your agents. It shows personal agents unless you name a team.

Personal key or team key

The key decides what the client may do. A personal key reaches agents in every team you belong to and lets the client run and edit them. A team connection works only after a team admin generates a team key; it sees only that team's agents and does not attribute actions to individual people. It is also narrower: 33 of the 87 tools refuse a team key, including start_agent, delete_agent, the trigger and evaluation tools, export_data, get_audit_logs, search_brain and resolve_session_approvals, and create_session fails as well. For daily building, pick personal.

What a connected client can do

The tool set covers agents, sessions, skills, artifacts, Brain search, evaluations and approvals. A client can create or update an agent, switch abilities such as web search, the browser or image generation, attach skills and MCP servers, add schedule or webhook triggers, and ask the Auto router which model it would pick. It can start, follow up and cancel sessions, and clear up to 20 pending approvals or human-input asks in one call. Evaluation runs accept a dry run that reports cost and skipped sessions before anything is queued. Member management answers only to organization Admins; custom-role credit limits and data exports are Enterprise features, and over MCP only Admins can use them.

The quieter feature is centralized auth. Through the same server, your client can call tools on Gmail, Slack, Linear and your custom MCP servers without setting up OAuth again in every client. The client receives tool results and never the apps' OAuth tokens, which stay in Gumloop. These calls follow the same governance as agents: custom roles, App Rules, domain restrictions, App Claims and Connector Activity logging. The MCP server and the REST API overlap heavily, but neither is a superset; queued messages, file uploads and chat completions are REST-only, covered in our Gumloop API guide.

Give your agents outside tools: custom, proxied or hosted

When a service has no built-in connector, Gumloop offers three ways to plug in an MCP server, and the real difference is who owns and administers it ^[2].

Custom MCP: a URL and a credential

A custom server needs a URL plus a credential stored for you alone or shared with your team. It must speak Streamable HTTP or SSE over a public HTTPS address. Authentication can be a Bearer token, one custom header, API key credentials or automatic OAuth discovery. One trap catches people: Anthropic models do not forward the custom header, so a server expecting something like X-API-Key needs the Bearer field instead, or an OpenAI, Gemini or Groq model. Gumball can also add a server from a pasted link, once you approve a card that shows its URL, name and visibility.

Proxied MCP: one registration for everyone

A proxied server is registered once for the whole organization. Gumloop logs every tool call and controls which tools people may use. Authentication is None, API Key or OAuth 2.0: the first two share one organization secret, while OAuth makes each person sign in with their own token. A server behind Cloudflare Access, AWS API Gateway or Azure API Management can use an Access Gate that stores the credential for that gate. Choose auth carefully. It cannot be changed after creation, and Fetch New Tools adds new tool names without updating existing definitions or removing old ones.

Hosted MCP: your code, Gumloop's servers

Hosted MCP suits a team that writes its own server. The code lives in a GitHub repository; Gumloop builds, deploys, monitors and routes requests, logging every tool call. Every push to the main branch triggers a new build, and an earlier deployment can be redeployed to roll back. The boilerplate is Python, but any language that implements MCP over HTTP works. Authentication is none, API key or credentials, or OAuth 2.0, fixed at creation. Pro includes one hosted server; Enterprise sets its own limit.

When a custom server will not connect

A server that works in Claude Desktop or Cursor can still fail in Gumloop ^[3]. The reason is location. Gumloop connects from its own cloud, so firewall rules, IP allowlists and OAuth redirect rules apply to Gumloop's addresses rather than your machine. Work through these checks in order, because the first ones explain most failures:

• Local server: localhost and STDIO are not supported. Expose the server through Cloudflare Tunnel or ngrok, or use Managed Tunnels on Enterprise.
• IP filtering: Gumloop's static egress IPs, which a firewall can allowlist, are an Enterprise feature; on Pro, drop the IP filter or rely on a Bearer token or OAuth instead. If the filter blocks Gumloop, a 403 hides the OAuth metadata and leaves OAuth 2.0 grayed out.
• OAuth: the provider must allow both Gumloop redirect URLs (the api.gumloop.com and api.gumstack.com callbacks) and support Dynamic Client Registration or Client ID Metadata Documents. Otherwise register your own OAuth app and choose OAuth 2.0 (Custom Client).
• Duplicates: if a teammate already added the same URL, your connection joins that server and keeps its name.
• The Gumloop server itself: a 403 on every tool, documentation search included, means your custom role has the Gumloop server switched off. An admin re-enables it on the role's Connectors tab.

The model matters too. On OpenAI and Anthropic models a custom server runs as native MCP, with the model provider connecting to it; Gemini and Groq models go through a backend connector where Gumloop executes the tool calls. When a server answers one model and not another, check that split before you rebuild anything.

What MCP calls cost, and who can see them

MCP tools bill like any connector. Every successful tool call costs at least 1 credit, and failed calls are free. With your own key for a tool such as Apollo or DataForSEO, the call drops to that 1-credit minimum. Model tokens, compute at 5 credits per session-minute of active processing and the 8% orchestration fee come on top. Gumloop's own examples put one run at 2 credits and another at 200, so read the chat details drawer before scaling; our buyers checklist covers the shared pool and the overage cap.

Watch Connector Activity. On Pro only Admins can open it, and it shows the last 24 hours; Enterprise keeps it with no history limit. Pro scopes connector policies per agent through tool approvals, while org-wide policies such as App Rules and App Claims, plus managed tunnels and static egress IPs, are Enterprise features. MCP servers also feed AI-created triggers, which can combine an external server such as Stripe's with built-in apps. Those triggers poll every 5 minutes by default and bill on every poll, fired or not; the Gumloop AI playbook explains how to set the polling interval sensibly.

Gumstack is the gateway, not the server

Gumstack is Gumloop's MCP gateway: a security and observability layer that logs every MCP tool call to a user, agent or service principal, controls which teams can use which MCP servers through RBAC and ABAC, handles secrets centrally, inventories MCP clients and servers across platforms rather than only Gumloop agents, and hosts custom MCP servers. Gumloop says it also flags unauthorized MCP servers and handles key rotation and revocation without local plaintext keys. It is sold through a demo, not self-serve, and Gumloop's about page lists the MCP gateway among Enterprise-tier features. Keep the names apart. The Gumloop MCP server comes with Pro.

Which MCP door to open first

If you live in Claude Code or Cursor and already build agents in Gumloop, start with the Gumloop MCP server and a personal key: it needs no plan beyond Pro and lets one assistant drive agents, approvals and connected apps without new OAuth setups. If your agents need a tool Gumloop lacks, add it as a custom server, then promote it to proxied once several teams depend on it and an admin wants logs and per-tool control. Write a hosted server only when no existing endpoint fits, and talk to Gumloop about Gumstack when the question becomes governance across every AI client, not one product.