Docs / Connectors
MCP
Give Claude Code, Codex or another MCP client access to your team's shared Pangea workspace.
The remote Pangea MCP server lets an MCP client hand work to your team's agents. It can provision servers and clusters, debug deployments, or answer questions about your cloud accounts. It is a tool interface into Pangea. It does not decide which runtimes Pangea can run.
This is different from a hosted runtime. With a hosted runtime, Pangea runs your Claude Code account on a workspace-isolated computer. With MCP, your own client calls Pangea remotely, and an agent on the team's configured runtime does the work.
Connect
The server speaks MCP Streamable HTTP, with JSON responses, at:
https://api.pangea.build/mcp
Claude Code
- Register the server:
bash
claude mcp add --transport http pangea https://api.pangea.build/mcp
- Open Claude Code. The server shows as Authorization needed and can't call tools yet.
- Run
/mcp, select pangea and choose Authenticate. Your browser opens the Pangea authorization page. - If you are already signed in to Pangea, click Authorize. Otherwise sign in first. Back in Claude Code, the server shows as connected.
Tokens refresh on their own after that, so you won't be asked again.
Codex
- Add the server to
~/.codex/config.toml:
toml
[mcp_servers.pangea]
url = "https://api.pangea.build/mcp"
- Run the OAuth flow from the terminal:
bash
codex mcp login pangea
Your browser opens the same Pangea authorization page.
Other MCP clients
Any client that supports Streamable HTTP with OAuth 2.1 discovery works the same way. The server's protected-resource metadata is at https://api.pangea.build/.well-known/oauth-protected-resource.
Clients without OAuth support can send a Pangea token directly:
Authorization: Bearer <your Pangea token>
Prefer OAuth where you can. A bearer token is a credential: keep it in a secure client config, never in a prompt or a repository.
In the Pangea app
Work started over MCP runs as real agents. Each one belongs to a team, runs on that team's connected runtime, and shows up in the workspace under the conversation's normal access controls.
What an external client can do
The server offers six tools. Your client lists them by their technical names.
| Tool | What it does |
|---|---|
| List teams | Your teams and their team_ids. Call this first. |
| Ask | Sends a prompt to the agent. Needs a team_id. Waits until the run finishes. |
| Check | Reconnects to an agent by session_id. Returns the answer, or running if the agent is still working. |
| List credentials | The cloud and tool credentials you can use in a team, and which are on for a conversation. |
| Update credentials | Replaces the set of credentials a conversation may use. |
| Capabilities | What the agent can do: supported task types and limits. |
In practice: while Codex builds your app, it can ask Pangea to "provision a Postgres on the staging cluster" or "find out why the deploy is failing". Pangea starts or resumes a shared agent, runs it on the team's runtime with the allowed accounts, and returns the result to Codex.
Conversations
Every conversation runs inside one of your teams, so pass a team_id from the list-teams tool to the ask tool.
Pass your own session_id to make a conversation resumable. If the ask call times out on the client side, the agent keeps running. Reconnect with the check tool and the same session_id to fetch the result. Reusing the session_id on a later ask continues the same conversation with its full history.
Permissions and authorization
MCP is an entry point, not an escape hatch. Agents started over MCP keep the same boundaries as agents started in the app.
- Who is acting. The OAuth flow signs in you, so every MCP call acts as your Pangea user with your access control: your team role and the access list on each connection. Pangea enforces this on every request. The agent can never use a credential your user isn't allowed to use.
- Credentials per conversation. A new conversation starts with every credential you can use in that team, except Resend, which sends email and must be turned on explicitly. To narrow it, call the update-credentials tool with the exact IDs to keep, taken from the list-credentials tool. It is a full replacement: any category you leave out or send empty is removed. It applies from the next ask in that conversation.
- What may run. Significant cloud work uses plans. Plans are not a gate on every write. The conversation's permission mode applies to each command the agent runs, as in the app.
Good to know
- The ask tool waits until the run finishes. For long tasks, pass a
session_idand poll with the check tool instead of relying on one long request. - Bearer-token access has exactly the same permissions as the OAuth flow. It is your user either way.