MCP client failing makes `opencode` hang.

[cvkmohan]

Description

Suppose we configure opencode with an MCP server - for example tidewave. Now tidewave will run well as long as phx.server is running fine. For whatever reason, suppose there is a compilation error and at that time, phx.server gets stopped - tidewave also gets stopped. Now, opencode which has in its config a running MCP tidewave server also stops working. I think it is not the desired behavior. If an MCP server fails to init - opencode should ignore it and move on - rather than becoming non-functional.

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

OpenCode version

No response

[github-actions[bot]]

This issue might be a duplicate of existing issues. Please check:

• #829: MCP server failed to start (unstable MCP servers) - Describes similar behavior where MCP connection failures during startup cause opencode to give up instead of gracefully handling the failure
• #1195: Some docker based MCPs prevent opencode shutdown - Shows how MCP server failures can prevent normal opencode operation and shutdown
• #504: big mess in terminal when MCP server fails - Documents poor error handling when MCP servers fail to initialize
• #1878: MCPClientError: Attempted to send a request from a closed client - Shows errors when MCP clients become unavailable during operation

Feel free to ignore if none of these address your specific case.

[rekram1-node]

I think dax addressed this in a version of opencode that ships next week

[venzen]

I have a related issue in v0.15.31 where the LLM becomes unresponsive when an MCP connection closes during a session:

8597 |   }
8598 |   onClose() {
8599 |     if (this.isClosed)
8600 |       return;
8601 |     this.isClosed = true;
8602 |     const error = new MCPClientError({
                         ^
MCPClientError: Connection closed
      cause: undefined,
 vercel.ai.error: true,
 vercel.ai.error.AI_MCPClientError: true,

      at onClose (../../node_modules/.bun/[email protected]+d6123d32214422cb/node_modules/ai/dist/index.mjs:8602:19)
      at close (../../node_modules/.bun/@[email protected]/node_modules/@modelcontextprotocol/sdk/dist/esm/client/sse.js:137:9)
      at close (../../node_modules/.bun/@[email protected]/node_modules/@modelcontextprotocol/sdk/dist/esm/client/sse.js:133:9)
      at close (../../node_modules/.bun/[email protected]+d6123d32214422cb/node_modules/ai/dist/index.mjs:8444:60)
      at close (../../node_modules/.bun/[email protected]+d6123d32214422cb/node_modules/ai/dist/index.mjs:8440:17)
      at map (1:11)
      at <anonymous> (src/mcp/index.ts:40:7)
      at <anonymous> (src/mcp/index.ts:38:11)

Is this expected behavior or shall I open a new bug for this?

[rekram1-node]

@venzen if u run opencode upgrade 1.0.65 I think this was fixex

[lovato]

I dug to find something similar to my problem, and I think this is it. On my company, the MCP server I am testing through OAuth needs both VPN and Proxy (not in prod yet). Fair. But if it happens for my laptop to just be disconnected from the VPN and I just need to do other stuff using opencode, it just hangs, since THAT MCP server isn't reachable right now. I could put helpers to inject/remove the config for this test phase, but maybe there is a better way to approach the problem. On VSCode, all MCP configuration are just configuration bits, and you turn on/off they as you need.