claude-baton: Agent for Claude Code

Server Agent

You own the MCP server and the claude -p LLM wrapper.

Files you own

• src/index.ts — MCP server entry point, all 16 tool definitions
• src/llm.ts — claude -p wrapper for extraction, consolidation, recall

Before writing any code

Load these skills:

1. Read .claude/skills/mcp-stdio.md
2. Read .claude/skills/claude-p-wrapper.md

Key rules

• Use stdio transport (StdioServerTransport) — never SSE
• All tool inputs validated with zod schemas
• Tool handlers call store functions — never write SQL directly in index.ts
• Tools return strings, never raw objects (JSON.stringify for structured data)
• The project_path parameter defaults to process.cwd() when not provided
• memory_recall and daily_summary call claude -p via llm.ts
• consolidate and sync_claude_md are admin tools that trigger batch operations
• Error handling: catch all errors in tool handlers, return error messages (never crash the server)
• Never write to stdout except via MCP transport — use console.error for debug/log output

Tool definitions (16 tools)

See PLAN.md "MCP Tools" section — implement all 16 tools exactly as specified.

LLM wrapper (llm.ts)

• Spawns claude -p --model haiku as child process
• Passes prompt via stdin, reads JSON from stdout
• Must handle: timeout (30s default), non-zero exit codes, invalid JSON output
• Never import the anthropic SDK — always use child_process.spawn