Request
Implement SEP-1686 (Tasks) in Claude Code's MCP client so long-running tool calls no longer block or hit the request timeout.
Why
MCP tool calls are synchronous and time out (DEFAULT_REQUEST_TIMEOUT_MSEC = 60000). Any tool whose work legitimately exceeds that (log searches, builds, bulk fetches, model inference) becomes unusable, and it's worse behind a proxy/gateway that adds its own edge timeout. SEP-1686 (accepted into the MCP spec 2025-11-25) is the standard fix: the server returns a task id and the client polls tasks/get / tasks/result instead of blocking.
Concrete case
A production Splunk MCP server sits behind a Cloudflare MCP portal. Async searches routinely run for minutes. Today the only options are a blocking tool call that risks the ~60s ceiling / edge timeout, or having the model poll on a cadence (one model turn per poll). Native task support would let the client poll the job over the protocol itself, reusing the existing MCP auth, with no side channel or CLI dependency.
Prior art
Previously requested in #52137 and #18617, both closed as not planned. Re-raising with a concrete production use case now that tasks have shipped in the spec.
Request
Implement SEP-1686 (Tasks) in Claude Code's MCP client so long-running tool calls no longer block or hit the request timeout.
Why
MCP tool calls are synchronous and time out (
DEFAULT_REQUEST_TIMEOUT_MSEC = 60000). Any tool whose work legitimately exceeds that (log searches, builds, bulk fetches, model inference) becomes unusable, and it's worse behind a proxy/gateway that adds its own edge timeout. SEP-1686 (accepted into the MCP spec 2025-11-25) is the standard fix: the server returns a task id and the client pollstasks/get/tasks/resultinstead of blocking.Concrete case
A production Splunk MCP server sits behind a Cloudflare MCP portal. Async searches routinely run for minutes. Today the only options are a blocking tool call that risks the ~60s ceiling / edge timeout, or having the model poll on a cadence (one model turn per poll). Native task support would let the client poll the job over the protocol itself, reusing the existing MCP auth, with no side channel or CLI dependency.
Prior art
Previously requested in #52137 and #18617, both closed as not planned. Re-raising with a concrete production use case now that tasks have shipped in the spec.