Can GLM-5.3 call Sume's hosted MCP? Function calling, not MCP

GLM-5.3 can call tools, and Sume's hosted MCP server has tools, but the model does not talk to the server by itself. Z.ai's GLM-5.3 page read on 2026-10-08 lists function calling and does not mention MCP. So a GLM agent needs an MCP client in your own code that lists Sume's tools, hands their schemas to the model as functions, and forwards each call to https://mcp.sume.com/mcp.

What the vendor page says

Do not read the missing mention as a statement that MCP cannot work. It means the page gives no client behaviour, so the bridge is yours.

The bridge, step by step

Sume's tools-and-gates page is clear that tools_list and tools_schema are the live contract, so generate the function definitions from them rather than copying a list.

• Open an MCP session to Sume with a Bearer token: an OAuth token with mcp:read and optionally mcp:write, or an API key.
• Call tools_list, then tools_schema for the tools you need, and convert each contract into a function definition.
• When the model returns a function call, send it as an MCP tools/call. Add idempotency_key on write and paid tools.
• Return the tool result to the model. For long jobs, call jobs_wait in slices of 45 to 55 seconds.

Or skip the bridge

If you do not want to run an MCP client, Agent Completions are plain HTTPS: POST /v1/agent/completions with an instruction and generation_spend_cap_usd, then poll /v1/agent-runs/{id}. The model there is sume-agent, so GLM is not involved, and Sume's agent calls its own tools. That route needs the agent_completions:write scope on the key.

Pick the bridge when you want GLM to do the reasoning, and Agent Completions when you want one call and a receipt. For the connector side of the same question, see Claude's MCP connector and Sume.

What to test first

Before you wire the whole loop, test three calls in order: mcp_health to confirm the credential and scopes, tools_list to confirm what the session sees, and one dry run of a paid tool. Each is cheap and tells you something about the bridge.

If the model returns arguments that do not match a contract, fetch the schema with tools_schema and show it to the model in the error message. Return the tool error text to the model along with the schema, and retry the same intent with the same idempotency_key.