Moth

Moth is a lightweight MCP server for project-local bug-fix analysis and verified fix memory.

What Moth Does

Moth receives error output through MCP, redacts likely secrets, normalizes the failure, detects the likely stack, checks project-local fix memory, and returns a structured fix brief.

Moth does not edit code, run shell commands, crawl repositories, require a backend, or maintain a global bug database.

Why Moth?

Bug-fix context is often local to a project: the command that failed, the framework in use, nearby configuration, and fixes that have already worked or failed in that repo.

Moth keeps that workflow small and explicit. It analyzes provided error context, suggests a best first fix, and records only verified fix outcomes in project-local memory.

Quick Start

Requires Node.js 18+.

Run directly:

npx -y @stfade/moth moth-mcp

Or install globally:

npm install -g @stfade/moth
moth-mcp

Generic MCP config

{
  "mcpServers": {
    "moth": {
      "command": "npx",
      "args": ["-y", "@stfade/moth", "moth-mcp"]
    }
  }
}

Usage Example

When using Moth with a supported AI agent, you can include a simple prompt like this along with your error:

"Use Moth to analyze this error before fixing it."

Supported Clients

| Client | Status | Setup | | --- | --- | --- | | Codex | Local plugin-ready | Setup | | Claude Code | Local plugin-ready | Setup | | Cursor | Plugin scaffold | Setup | | Gemini CLI | Extension scaffold | Setup | | Gemini Antigravity | MCP config-ready | Setup | | OpenCode | MCP config-ready | Setup | | Generic MCP | Config-ready | Setup |

“Local plugin-ready” means the integration wrapper is included and can be tested locally. Marketplace submission and approval are not included yet.

Tools

Moth exposes exactly two MCP tools.

analyze_error

Analyzes provided error output before a fix is attempted.

Input fields:

• error_output
• command?
• cwd?
• package_context?
• relevant_files?
• environment?

Output fields:

• analysis_id
• fingerprint
• stack
• likely_cause
• best_first_fix
• verification
• prior_project_fixes
• avoid
• confidence

remember_fix_result

Records verified project-local fix memory.

Input fields:

• analysis_id
• fingerprint
• stack
• fix_attempted
• verification_command
• verification_result: "passed" | "failed"
• notes?

The public worked input is rejected. worked is derived from verification_result.

Verified Memory Lifecycle

analyze_error
→ apply/attempt fix
→ run verification command
→ remember_fix_result

Call remember_fix_result only when:

1. a fix/change was actually attempted
2. the verification command actually ran
3. the result is clearly passed or failed

Do not call it for suggestions, skipped changes, missing verification, ambiguous results, or guesses.

Local Memory

Verified project-local fix memory is stored at:

.moth/fix-memory.jsonl

Moth keeps a small Moth-owned analysis registry outside the project so remember_fix_result can map analysis_id back to the correct project path after an MCP server restart.

Skills

Moth includes concise skills for compatible agents:

• moth-debug-first-fix
• moth-source-backed-research
• moth-verify-fix

The MCP server itself does not perform live web research. Compatible agents may use their own search tools, guided by Moth skills, when external sources are needed.

Safety

• read-only by default
• no source edits
• no shell execution
• no repo-wide scan
• no background watcher
• no external service required
• redacts likely secrets before analysis, responses, and memory writes

Development

pnpm install
pnpm test
pnpm build
pnpm dev
npm pack --dry-run

License

MIT