OpenCode failed with ‘ProviderModelNotFoundError.’ The model ID existed—just not spelled the way my config spelled it.

My MCP server config worked in Claude Desktop. It failed validation in OpenCode. Same protocol, different schemas.

Stop invoking AI skills manually. Configure them to auto-load, turning one-off commands into persistent background behaviors.

How a simple naming mismatch in an example file during a project rebrand led to a critical authentication failure.

How assuming shell-like tilde expansion in a JSON config led to a plugin loading failure and what it taught me about path handling.

How to configure skills to auto-load for specific AI agents, turning one-off invocations into persistent behaviors.

MCP config that works in Claude Desktop fails in OpenCode. Same protocol, different schema. A tale of subtle incompatibilities.

The Idea

Configure skills to auto-load for specific agents via oh-my-opencode.json.

Context

Wanted content-seed harvesting to be always active, not requiring manual /content-seed invocation each session.

Solution: Add skills array to agent config.

Implementation

// ~/.config/opencode/oh-my-opencode.json
{
  "agents": {
    "sisyphus": {
      "model": "...",
      "skills": ["content-seed"]  // ← auto-loads
    }
  }
}

Why It Matters

• Skills become persistent behaviors, not one-off invocations
• Enables “always-on” passive agents (harvesters, monitors, etc.)
• Configuration as behavior definition

Content Angle

• Tutorial: “Making OpenCode skills always-active”
• Pattern: “Passive AI behaviors through skill injection”

What Happened

OpenCode config validation failed with “Invalid input mcp.mcp-image” error. The MCP server configuration used wrong schema format.

The Broken Config

"mcp-image": {
  "type": "stdio",           // WRONG
  "command": "npx",          // WRONG
  "args": ["-y", "mcp-image"], // WRONG
  "env": {                   // WRONG field name
    "GEMINI_API_KEY": "...",
    "IMAGE_OUTPUT_DIR": "..."
  }
}

Investigation

1. First tried adding "type": "stdio" - still failed
2. Fetched official OpenCode docs at opencode.ai/docs/mcp-servers
3. Discovered correct schema from docs

Resolution

"mcp-image": {
  "type": "local",                           // local, not stdio
  "command": ["npx", "-y", "mcp-image"],    // array, not separate args
  "environment": {                           // environment, not env
    "GEMINI_API_KEY": "...",
    "IMAGE_OUTPUT_DIR": "..."
  }
}

Key differences:

• type: "local" not "stdio" (OpenCode terminology)
• command is a single array including args, no separate args field
• environment not env for env vars

Lesson

OpenCode MCP schema differs from Claude Desktop’s MCP schema. Always check vendor docs - don’t assume schema compatibility between different MCP hosts.

What Happened

ProviderModelNotFoundError when starting OpenCode. The oh-my-opencode.json config referenced model IDs that don’t exist in the provider definitions.

The Error

ProviderModelNotFoundError
providerID: "google"
modelID: "gemini-3-flash"
suggestions: ["gemini-3-flash-preview", "antigravity-gemini-3-flash"]

Investigation

1. Error showed suggestions for valid model names
2. Searched config files for the invalid model ID
3. Found references in ~/.config/opencode/oh-my-opencode.json:
   • Line 19: multimodal-looker agent
   • Line 39: visual-engineering category
   • Line 46: artistry category
   • Line 60: writing category

Resolution

Changed all occurrences to use valid model IDs:

// Before (invalid)
"model": "google/gemini-3-flash"
"model": "google/gemini-3-pro"

// After (valid - using antigravity provider)
"model": "google/antigravity-gemini-3-flash"
"model": "google/antigravity-gemini-3-pro"

Lesson

Model IDs in oh-my-opencode.json must exactly match the model IDs defined in opencode.json provider config. The error’s suggestions field shows valid alternatives - use those exact strings.