Troth is a structural map of the codebase. The tree of specs and sub-specs IS the architecture of the system Troth is tracking. They evolve together.
Not a project tracker. Not a wiki. Not a kanban board. Not a sprint planner.
https://mcp.troth.dev/sse Add as a remote MCP server in Claude.ai, Claude Code, Cursor, or any MCP-compatible agent.
npx troth-dev init Add the Troth MCP server to Claude.ai, Claude Code, or any MCP-compatible agent. Your agent gets structured project context automatically.
Describe what you want built. Troth decomposes it into discrete tasks with acceptance criteria and bounded file scope.
Each task must pass verification gates before completion. The agent that writes the code doesn't verify it — Troth does.
A bootstrap run generates structured YAML that agents use as ground truth.
name: auth
description: JWT authentication with refresh token rotation
architecture:
pattern: stateless JWT + Redis refresh store
entry_points:
- src/auth/middleware.ts
- src/auth/routes.ts
tasks:
- id: auth-middleware
objective: Validate JWT on every protected route
acceptance_criteria:
- 401 on missing token
- 401 on expired token
- 403 on insufficient scope
verification:
- command: bun test tests/auth.test.ts
Specs decompose into tasks with dependencies. Troth surfaces what's ready and keeps parallel work conflict-free.
Every run, every decision, every file touched — logged and reviewable. Requests link work to the reasons it exists.
Each task declares exactly which files it may touch. Agents don't wander.
Verification gates must pass before a task is complete. The agent writing code can't mark its own homework.