Continue: add Sume hosted MCP, agent mode only, timeout above 55 s

Continue runs MCP tools in agent mode only. Add Sume's hosted MCP as a streamable-http YAML file, send a key header, and raise requestOptions timeout past 55 s.

5 min readSume
All posts

Yes, Continue can call Sume's hosted MCP server, but only from its agent mode: Continue's own MCP page says MCP "can only be used in the agent mode", so chat and plan sessions will not see Sume's tools. The setup is one YAML file with type: streamable-http and the URL https://mcp.sume.com/mcp, plus a key header and a request timeout above the 55 seconds that one Sume jobs_wait call can hold.

Continue's side of this comes from its MCP deep dive and its config.yaml reference, both read on 2026-10-11. Sume's side comes from the MCP quickstart, OAuth and API keys and Jobs and results. Sume has no Continue-specific integration; Continue connects to the same remote endpoint as any other streamable HTTP client.

Where does the Sume entry go?

Continue reads MCP servers either from an mcpServers block in the main config or from separate files in a .continue/mcpServers folder at the top of your workspace. It also accepts JSON MCP files copied from other tools into that folder. A separate file keeps the Sume entry out of the config you share, which matters once a key is involved.

For a remote server the page gives three required keys: name, type set to streamable-http, and url. The reference page lists requestOptions as the optional block for sse and streamable-http servers, and connectionTimeout as the timeout for the initial connection.

# .continue/mcpServers/sume.yaml
mcpServers:
  - name: sume
    type: streamable-http
    url: https://mcp.sume.com/mcp
    requestOptions:
      timeout: 90000
      headers:
        x-api-key: ${{ secrets.SUME_API_KEY }}

What does each field do for Sume?

The table maps each Continue field to the Sume behavior it has to match. Continue's field descriptions are from the reference page read on 2026-10-11; Sume's are from its docs.

Continue fields against Sume hosted MCP rules (Continue pages read 2026-10-11)
Continue fieldWhat Continue saysWhat Sume needs
type: streamable-httpRemote transport for cloud-hosted serversSume's endpoint is a remote HTTP MCP at mcp.sume.com/mcp
requestOptions.headersCustom HTTP headersAuthorization: Bearer or x-api-key with a Sume key
requestOptions.timeoutTimeout for requestsAbove 55 s, because one jobs_wait call can hold 55 s
connectionTimeoutTimeout for the initial connectionOnly the first connect; not the tool-call hold
Agent modeThe only mode where MCP worksNeeded before tools_list shows anything

Which auth should Continue use with Sume?

Sume's docs describe two credentials for hosted MCP. OAuth is the path for interactive clients, and read-only by default: mcp:read sees read-only tools, and the Write toggle on the consent page adds mcp:write. An API key sees the full tool set, and paid or write calls still need an idempotency_key. The two are not interchangeable: an OAuth token is not an API key.

Continue's pages I read name requestOptions with headers but describe no OAuth sign-in for MCP servers, so the header route is the one the documentation supports. Create a key in the dashboard, export it as SUME_API_KEY, and keep it out of the YAML. Continue's page shows the ${{ secrets.NAME }} pattern for an env value; whether that expansion also applies inside requestOptions.headers is not stated in the pages I read, so check the first call: a 401 means the header never arrived with a real key.

Why raise the timeout past 55 seconds?

Sume's jobs_wait tool holds one HTTP request open for up to 55 seconds (default 50) and then returns a status snapshot. Sume's docs say a longer hold would die at the edge, so a ten-minute render is several waits on the same job ids, never one long wait. If Continue's request timeout is shorter than the hold, the editor can give up on a call while the job keeps running and billing.

Set requestOptions.timeout above 55 seconds. The reference page does not state the unit, so confirm it on your build; the example above uses 90000 on the assumption of milliseconds. When a wait returns wait_slice_expired, call jobs_wait again with the same ids. Never repeat the paid create call, and send a fresh idempotency_key only for a genuinely new job.

How do I check that it works?

Switch Continue to agent mode and ask it to call tools_list, then mcp_health. Sume's quickstart lists both as read-only first calls, and mcp_health reports the auth source. With an API key the full set appears; a read-only OAuth session would hide the generation tools.

Before the first paid call, ask for a dry_run on generate_image or generate_video. Sume's docs recommend dry_run=true or generation_admission_preview before the first paid submit, and the optional max_spend_usd caps spend only when you send it. Continue's page I read documents no tool-approval behavior, so the key's wallet and those arguments are your guardrails, not the editor.

Sources

Related posts

More in Integrations

All Integrations posts

Written by Sume