Skip to content

opencode-mcp.md

Latest commit

 

History

History
137 lines (104 loc) · 3.08 KB

opencode-mcp.md

File metadata and controls

137 lines (104 loc) · 3.08 KB

OpenCode MCP Integration

This guide connects OpenCode to Code Agent++ through the stdio MCP server. OpenCode remains the coding runtime; Code Agent++ supplies task context, edit boundaries, evidence checks, and repair/finalize decisions.

Configuration

Build the package:

npm run build

Add the MCP server to the OpenCode MCP configuration supported by your OpenCode version:

{
  "mcpServers": {
    "code-agent-plusplus": {
      "command": "code-agent-plusplus-mcp",
      "args": []
    }
  }
}

Local development can use the built server directly:

{
  "mcpServers": {
    "code-agent-plusplus": {
      "command": "node",
      "args": ["F:/codex/code/Repo-to-Agent-Context/dist/mcp/server.js"]
    }
  }
}

Start Loop

Ask OpenCode to call:

code_agent_plusplus_start_loop

Example:

{
  "repo": "F:/path/to/repo",
  "task": "fix login timeout bug",
  "agent": "other",
  "type": "bugfix",
  "base": "main"
}

The response is the runtime contract for the turn: nextAction, blocking, requiredCommands, mustInspect, allowedEditGlobs, avoidEditGlobs, and missingEvidence.

Step

Record file edits:

{
  "repo": "F:/path/to/repo",
  "traceId": "fix-login-timeout-bug",
  "agent": "opencode",
  "action": "edit",
  "files": ["src/auth/session.ts"],
  "reason": "Timeout handling lives in the session module"
}

Record command evidence:

{
  "repo": "F:/path/to/repo",
  "traceId": "fix-login-timeout-bug",
  "agent": "opencode",
  "action": "run-test",
  "command": "npm test -- test/auth/session.test.ts",
  "result": "passed"
}

If OpenCode emits JSON events, prefer passing those events through the CLI orchestrator or future native normalizer; MCP step is the portable manual bridge.

Evaluate

Call:

code_agent_plusplus_evaluate

Example:

{
  "repo": "F:/path/to/repo",
  "task": "fix login timeout bug",
  "traceId": "fix-login-timeout-bug",
  "base": "main",
  "phase": "after-edit",
  "failOn": "required"
}

OpenCode should treat blocking: true as a stop condition and follow nextAction plus requiredCommands.

Repair

Call code_agent_plusplus_repair when evaluation returns missing evidence, policy failures, contract failures, hallucination findings, or regression findings. Repair should focus on the returned requiredActions and must not widen the edit surface unless allowedEditGlobs changes after a repack.

Finalize

Call code_agent_plusplus_finalize only when the latest evaluation is non-blocking:

{
  "repo": "F:/path/to/repo",
  "task": "fix login timeout bug",
  "traceId": "fix-login-timeout-bug",
  "base": "main",
  "finalState": "success"
}

Limitations

  • MCP mode is advisory unless your OpenCode workflow refuses to continue on blocking: true.
  • OpenCode event schemas may differ by version; native event normalization should be validated per version.
  • For Code Agent++-led execution, prefer code-agent-plusplus orchestrate ... --executor opencode.
  • MCP step does not execute shell commands; it records evidence provided by the agent or host.