Docs

MCP overview

The Seamward MCP server can give an AI client read-only access to one workspace mode and a reviewed connection set. The server enforces that selection on every request. Browser consent binds the selected Test or Live mode and exact connections before approval.

Invitation-only access

The Seamward MCP server is available to invited workspaces. Your workspace's MCP server URL is provided during onboarding, and the interface may change before broader access. See current availability for the supported connection and service boundaries.

What you can do

After connecting, ask your client questions like these:

  • "Which of our integrations have open incidents right now?"
  • "Summarise the newest incident and the evidence behind it."
  • "Did replay validation pass for the payout-webhook repair proposal?"
  • "Fetch the approved repair bundle for proposal rep_7f3k2m and prepare a local patch from it."

The client answers by calling Seamward's read-only tools. The resources and tools reference lists all eight tools with their parameters and example results.

Requirements

  • A Seamward account with membership in the workspace you want to connect. Your role does not matter for connecting; every grant is read-only.
  • Your MCP server URL from workspace onboarding. It always ends in /mcp, for example https://mcp.seamward.example/mcp.
  • An MCP client that supports the streamable HTTP transport and OAuth. Claude Code, Claude (web and desktop), and Cursor all qualify, as does any other client with the same support.

This hosted workspace MCP does not need an API key, ingest token, or any other manually copied credential. Authorization happens entirely through your browser sign-in. Never paste a Seamward ingest token or browser cookie into this hosted MCP client; the server does not accept them.

The local coding-agent setup MCP is a different server. It changes and verifies one codebase and uses its own browser authorization bound to one exact workspace environment. It previews the exact Integrations it will create or reuse before any remote write. See automatic setup for the one-command project installation.

Client configuration

The following client settings start the reviewed browser consent flow. Invitation access and service boundaries are described in current availability.

Replace https://mcp.seamward.example/mcp in each example with the MCP server URL provided during workspace onboarding.

Claude Code

Add the server from your terminal:

Code example

claude mcp add --transport http seamward https://mcp.seamward.example/mcp

Then run /mcp inside Claude Code and choose seamward to start authentication. Your browser opens the Seamward approval flow described below.

Claude (web and desktop)

Open Settings, then Connectors, then choose Add custom connector. Name it Seamward, paste your MCP server URL, and save. Claude starts the browser approval flow the first time you use the connector.

Cursor

Add the server to .cursor/mcp.json in your project (or ~/.cursor/mcp.json for all projects):

Code example

{  "mcpServers": {    "seamward": {      "url": "https://mcp.seamward.example/mcp"    }  }}

Cursor shows a Needs login state for the server; click it to start the browser approval flow.

Any other MCP client with streamable HTTP and OAuth support works the same way: give it the server URL, and it discovers the authorization details automatically.

Approve access

Approve access through this browser flow:

  1. Your browser opens the Seamward sign-in page. Sign in with your verified account.
  2. Seamward shows the consent screen. Choose one workspace, one mode, and the exact connections this client may read. An empty selection grants no connection evidence.
  3. Review the requested read scopes. The authorization page explains exactly what each scope unlocks.
  4. Choose Allow only after the screen confirms that exact selection. The client then receives a short-lived access token bound to it.

Access tokens expire after 15 minutes. Clients that request offline access refresh automatically in the background; otherwise the client asks you to approve again when the token expires. No long-lived secret is ever stored in the client.

Verify the connection

Ask the client for a workspace overview, for example: "Give me an overview of my Seamward workspace." The client calls the get_workspace_overview tool and receives a result like this:

Code example

{  "schemaVersion": "seamward.mcp/1",  "workspace": {    "id": "wks_2ac9017e",    "name": "Acme",    "slug": "acme"  },  "data": {    "tenant": { "id": "ten_4b7e2f81", "name": "Acme" },    "integrations": 4,    "incidents": { "open": 1, "total": 7 },    "latestObservationAt": "2026-08-14T18:22:31.000Z"  }}

The connection also appears under Workspace settings → MCP with its reviewed mode, connection set, scopes and connection date. That page is also where you revoke access. A changed mode or connection set requires a new review; a removed connection invalidates the previous selection.

Troubleshooting

The client reports unauthorized or a 401. The access token has expired and the client has no refresh token, or the client's access was revoked in Workspace settings → MCP. Reconnect the client and approve again.

A tool call returns forbidden. The grant does not include the scope that tool needs. This is most common with get_approved_repair_bundle, because the export scope is not part of the default request. Reconnect and approve a request that includes seamward:repairs:export.

A tool call returns not_found. The id is outside the reviewed workspace, mode, or connection set, or no longer exists. The server does not reveal which case applies. Check the reviewed selection and renew authorization if the client needs another mode or connection.

The server does not appear in the client. Check the URL is exactly the one from onboarding, including the /mcp path, with no trailing slash. Opening the URL in a normal browser tab returns a method_not_allowed error; that is expected and confirms the server is reachable, because the server only accepts POST /mcp requests from MCP clients.

Analysis health

Integration list and detail tools include the same structuralAnalysis and behavioralAnalysis state as the Workspace API. A connection with observations and no open incidents can still have structural or behavioral analysis paused. Check status, lastSuccessAt, and the capacity code before describing a connection as healthy. See structural capacity recovery and behavioral capacity recovery for limits and recovery behavior.

Next steps