Find waste and bad practices in AI agents.

Local-first telemetry and linting for chats, MCP servers, skills, tools, and model/resource usage.

It records structural metadata by default—not prompts, tool arguments, or tool results—and turns traces into actionable waste findings.

Node 24+ is recommended. Node 22.13+ is supported but may emit a SQLite experimental warning.

npm install
npm run build
npm link
agenttidy serve

Open http://127.0.0.1:4318.

Wrap an MCP stdio server:

node dist/src/cli.js mcp --server filesystem --collector http://127.0.0.1:4318 -- npx -y @modelcontextprotocol/server-filesystem .

Connect to a Streamable HTTP MCP server (add --header 'Authorization: ******' for authenticated servers):

node dist/src/cli.js mcp --server remote --url https://mcp.example.com/mcp --collector http://127.0.0.1:4318

The HTTP mode bridges the agent's stdio MCP connection to the remote endpoint and supports JSON or SSE responses.

Get a terminal report:

node dist/src/cli.js report

See GETTING_STARTED.md for chat/skill instrumentation.

• MCP001 mostly unused tool exposure
• MCP002 oversized tool schema
• MCP003 duplicate invocation
• MCP004 large tool result
• MCP005 repeated failed call
• CHAT001 history dominates current turn
• CHAT002 likely correction tax
• SKILL001 loaded but unused skill
• SKILL003 large skill context

Token sizes derived from byte size are explicitly estimates. Provider-reported token counts remain the source of truth when available.