OpenCode MCP and token overhead

Each enabled MCP server in OpenCode adds tool schemas to agent context on every turn—even when no MCP tool is called. That is the main MCP token overhead.

This page answers opencode disable mcp, lazy-load patterns, and when to prefer built-in grep / glob.

Hub: Tokenminning in OpenCode.

Why MCP blows up context

What loads	When	Cost
Tool names + JSON schemas	Every agent turn	Input tokens
Tool results	When invoked	Input on later turns

OpenCode MCP docs warn that GitHub MCP and similar broad servers can exceed context limits quickly. Five overlapping servers can dwarf your actual prompt.

Measure before/after with opencode stats when enabling a new server.

Disable MCP servers

Turn off whole servers in opencode.json:


{
  "mcp": {
    "github": { "enabled": false },
    "playwright": { "enabled": false }
  }
}

Disable MCP tools globally, enable per-agent when needed:


{
  "tools": { "my-mcp*": false }
}

Per-agent patterns (enable only on Build):


{
  "agent": {
    "plan": {
      "tools": { "github-mcp*": false }
    }
  }
}

Check your OpenCode version’s agent tool permissions for exact syntax.

Lazy-load pattern

Lazy MCP means: default off, enable only for the agent and week you need it.

Set "enabled": false on all non-essential servers
Enable one narrow server for the current task (e.g. Playwright only while debugging UI)
Run opencode stats --days 1 before and after

Skills already lazy-load full content; MCP does not — schemas are always present until disabled.

Built-in tools vs MCP

For local codebases, prefer:

Need	Use	Avoid
Text search	Built-in `grep` / `glob`	MCP codebase search
Read file	Built-in `read`	Pasting file in chat
GitHub API	MCP only when remote PR/issue data required	MCP for local git status

One task-specific MCP server beats five general-purpose integrations.

Skills vs MCP

Skills load metadata every turn but full content only when the skill tool runs. Good for occasional workflows.

Disable skill tool on Plan if you want zero skill metadata in planning sessions
Use permission.skill to hide experimental skills

Troubleshooting

Spike after enabling MCP — Schemas attach every turn. Disable globally; enable per-agent.

High input, low tool use — MCP bloat without invocations. Audit enabled servers.

Context overflow with MCP — Combine disable MCP with compaction.