[feature request]: ability to dynamic turn on mcp server for specific agent

[zelifish]

Hope I'm not doing this wrong, but with this opencode config in version 0.11.1

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "graphql": {
      "type": "local",
      "enabled": false,
      "command": ["bunx", "mcp-graphql"],
      "environment": {
        "ENDPOINT": "http://localhost:13087/graphql"
      }
    }
  }
}

And this agent definition

---
description: test agaent
mode: primary
tools:
  graphql_*: true
---

This is a test agent

My agent can't actually see the graphql mcp tools. I tried various permutations of graphql, graphql*, graphql_*, etc.

[github-actions]

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

• Feature Request: Global default tools configuration for agents #2888: Discussion about global default tools configuration for agents and tool inheritance behavior
• Enable/Disable MCP server by Mode #1101: Request for enabling/disabling MCP servers by mode, with detailed tool configuration syntax
• Ability to restrict MCP tool use to subagent only #2277: Request to restrict MCP tool use to specific agents/subagents only
• Option to Enable / Disable MCP Servers (via TUI) #514: Feature request for enabling/disabling MCP servers at runtime

Feel free to ignore if none of these address your specific case with disabled MCP servers not being accessible through agent tool configurations.

[rekram1-node]

@zelifish so here is what you need to do:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "graphql": {
      "type": "local",
      "enabled": true,
      "command": ["bunx", "mcp-graphql"],
      "environment": {
        "ENDPOINT": "http://localhost:13087/graphql"
      }
    }
  },
  "tools": {
    "graphql*": false
  }
}

enabled being false means the MCP server won't ever turn on at all so it isn't possible to give to llm

[zelifish]

Hmm, my motivation for this was to mitigate long startup times for tools that we only need when working on the front end portion of our repo. Is there a better way of doing that? My only other solution was to move my opencode configs down into the FE section but then I can't actually use them when working across the back and the front end having started opencode from a higher directory (or can I)?

I guess I could explicitly pass the config in the opencode command but no one's going to obey my complicated instructions if that's what I make them do.

[rekram1-node]

@zelifish rn there isn't a better way of doing it tbh, in opentui version there will probably be a mechanism to enable and disable mcp servers on the fly but rn it isn't super easy

[benceferdinandy-signifyd]

@rekram1-node do you have an estimate for opentui release?

[rekram1-node]

@benceferdinandy-signifyd if you need it specifically for this issue, i can't really give a timeline on when this will be addressed as it is a lower priority rn

As for opentui version shipping in general id say next week probs

[benceferdinandy-signifyd]

Just generally opentui, thanks!

[emilford]

@rekram1-node the suggested approach doesn't seem to work for me. I have the GitHub MCP server set up and enabled, tools in the global configuration includes "github*": false and the subagent markdown includes tools: "github*": true, but the agent doesn't have access to the tools. If I comment out the global tools line everything works. What am I missing? 🤔

Edit:

Ah, it looks like configuring the agent tools in opencode.js works in the way described above. Is this a markdown limitation?

  "tools": {
    "github*": false,
    ...
  },
  "agent": {
    "github": {
      "tools": {
        "github*": true
      }
    }
  }